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 "keylist.h"
  20 #include "move.h"
  21 #include "replicas.h"
  22 #include "super-io.h"
  23 #include "trace.h"
  24
  25 #include <linux/ioprio.h>
  26 #include <linux/kthread.h>
  27
  28 static void trace_move_extent2(struct bch_fs *c, struct bkey_s_c k)
  29 {
  30         if (trace_move_extent_enabled()) {
  31                 struct printbuf buf = PRINTBUF;
  32
  33                 bch2_bkey_val_to_text(&buf, c, k);
  34                 trace_move_extent(c, buf.buf);
  35                 printbuf_exit(&buf);
  36         }
  37 }
  38
  39 static void trace_move_extent_read2(struct bch_fs *c, struct bkey_s_c k)
  40 {
  41         if (trace_move_extent_read_enabled()) {
  42                 struct printbuf buf = PRINTBUF;
  43
  44                 bch2_bkey_val_to_text(&buf, c, k);
  45                 trace_move_extent_read(c, buf.buf);
  46                 printbuf_exit(&buf);
  47         }
  48 }
  49
  50 static void trace_move_extent_alloc_mem_fail2(struct bch_fs *c, struct bkey_s_c k)
  51 {
  52         if (trace_move_extent_alloc_mem_fail_enabled()) {
  53                 struct printbuf buf = PRINTBUF;
  54
  55                 bch2_bkey_val_to_text(&buf, c, k);
  56                 trace_move_extent_alloc_mem_fail(c, buf.buf);
  57                 printbuf_exit(&buf);
  58         }
  59 }
  60
  61 static void progress_list_add(struct bch_fs *c, struct bch_move_stats *stats)
  62 {
  63         mutex_lock(&c->data_progress_lock);
  64         list_add(&stats->list, &c->data_progress_list);
  65         mutex_unlock(&c->data_progress_lock);
  66 }
  67
  68 static void progress_list_del(struct bch_fs *c, struct bch_move_stats *stats)
  69 {
  70         mutex_lock(&c->data_progress_lock);
  71         list_del(&stats->list);
  72         mutex_unlock(&c->data_progress_lock);
  73 }
  74
  75 struct moving_io {
  76         struct list_head                read_list;
  77         struct list_head                io_list;
  78         struct move_bucket_in_flight    *b;
  79         struct closure                  cl;
  80         bool                            read_completed;
  81
  82         unsigned                        read_sectors;
  83         unsigned                        write_sectors;
  84
  85         struct bch_read_bio             rbio;
  86
  87         struct data_update              write;
  88         /* Must be last since it is variable size */
  89         struct bio_vec                  bi_inline_vecs[0];
  90 };
  91
  92 static void move_free(struct moving_io *io)
  93 {
  94         struct moving_context *ctxt = io->write.ctxt;
  95
  96         if (io->b)
  97                 atomic_dec(&io->b->count);
  98
  99         bch2_data_update_exit(&io->write);
 100
 101         mutex_lock(&ctxt->lock);
 102         list_del(&io->io_list);
 103         wake_up(&ctxt->wait);
 104         mutex_unlock(&ctxt->lock);
 105
 106         kfree(io);
 107 }
 108
 109 static void move_write_done(struct bch_write_op *op)
 110 {
 111         struct moving_io *io = container_of(op, struct moving_io, write.op);
 112         struct moving_context *ctxt = io->write.ctxt;
 113
 114         if (io->write.op.error)
 115                 ctxt->write_error = true;
 116
 117         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 118         atomic_dec(&io->write.ctxt->write_ios);
 119         move_free(io);
 120         closure_put(&ctxt->cl);
 121 }
 122
 123 static void move_write(struct moving_io *io)
 124 {
 125         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 126                 move_free(io);
 127                 return;
 128         }
 129
 130         closure_get(&io->write.ctxt->cl);
 131         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 132         atomic_inc(&io->write.ctxt->write_ios);
 133
 134         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
 135 }
 136
 137 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
 138 {
 139         struct moving_io *io =
 140                 list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
 141
 142         return io && io->read_completed ? io : NULL;
 143 }
 144
 145 static void move_read_endio(struct bio *bio)
 146 {
 147         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 148         struct moving_context *ctxt = io->write.ctxt;
 149
 150         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 151         atomic_dec(&ctxt->read_ios);
 152         io->read_completed = true;
 153
 154         wake_up(&ctxt->wait);
 155         closure_put(&ctxt->cl);
 156 }
 157
 158 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt,
 159                                         struct btree_trans *trans)
 160 {
 161         struct moving_io *io;
 162
 163         if (trans)
 164                 bch2_trans_unlock(trans);
 165
 166         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 167                 list_del(&io->read_list);
 168                 move_write(io);
 169         }
 170 }
 171
 172 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt,
 173                                        struct btree_trans *trans)
 174 {
 175         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 176
 177         move_ctxt_wait_event(ctxt, trans,
 178                 !atomic_read(&ctxt->write_sectors) ||
 179                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 180 }
 181
 182 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 183 {
 184         struct bch_fs *c = ctxt->c;
 185
 186         move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 187         closure_sync(&ctxt->cl);
 188
 189         EBUG_ON(atomic_read(&ctxt->write_sectors));
 190         EBUG_ON(atomic_read(&ctxt->write_ios));
 191         EBUG_ON(atomic_read(&ctxt->read_sectors));
 192         EBUG_ON(atomic_read(&ctxt->read_ios));
 193
 194         if (ctxt->stats) {
 195                 progress_list_del(c, ctxt->stats);
 196                 trace_move_data(c,
 197                                 atomic64_read(&ctxt->stats->sectors_moved),
 198                                 atomic64_read(&ctxt->stats->keys_moved));
 199         }
 200
 201         mutex_lock(&c->moving_context_lock);
 202         list_del(&ctxt->list);
 203         mutex_unlock(&c->moving_context_lock);
 204 }
 205
 206 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 207                            struct bch_fs *c,
 208                            struct bch_ratelimit *rate,
 209                            struct bch_move_stats *stats,
 210                            struct write_point_specifier wp,
 211                            bool wait_on_copygc)
 212 {
 213         memset(ctxt, 0, sizeof(*ctxt));
 214
 215         ctxt->c         = c;
 216         ctxt->fn        = (void *) _RET_IP_;
 217         ctxt->rate      = rate;
 218         ctxt->stats     = stats;
 219         ctxt->wp        = wp;
 220         ctxt->wait_on_copygc = wait_on_copygc;
 221
 222         closure_init_stack(&ctxt->cl);
 223
 224         mutex_init(&ctxt->lock);
 225         INIT_LIST_HEAD(&ctxt->reads);
 226         INIT_LIST_HEAD(&ctxt->ios);
 227         init_waitqueue_head(&ctxt->wait);
 228
 229         mutex_lock(&c->moving_context_lock);
 230         list_add(&ctxt->list, &c->moving_context_list);
 231         mutex_unlock(&c->moving_context_lock);
 232
 233         if (stats) {
 234                 progress_list_add(c, stats);
 235                 stats->data_type = BCH_DATA_user;
 236         }
 237 }
 238
 239 void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
 240 {
 241         memset(stats, 0, sizeof(*stats));
 242         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 243 }
 244
 245 static int bch2_extent_drop_ptrs(struct btree_trans *trans,
 246                                  struct btree_iter *iter,
 247                                  struct bkey_s_c k,
 248                                  struct data_update_opts data_opts)
 249 {
 250         struct bch_fs *c = trans->c;
 251         struct bkey_i *n;
 252         int ret;
 253
 254         n = bch2_bkey_make_mut_noupdate(trans, k);
 255         ret = PTR_ERR_OR_ZERO(n);
 256         if (ret)
 257                 return ret;
 258
 259         while (data_opts.kill_ptrs) {
 260                 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
 261                 struct bch_extent_ptr *ptr;
 262
 263                 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
 264                 data_opts.kill_ptrs ^= 1U << drop;
 265         }
 266
 267         /*
 268          * If the new extent no longer has any pointers, bch2_extent_normalize()
 269          * will do the appropriate thing with it (turning it into a
 270          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 271          */
 272         bch2_extent_normalize(c, bkey_i_to_s(n));
 273
 274         /*
 275          * Since we're not inserting through an extent iterator
 276          * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
 277          * we aren't using the extent overwrite path to delete, we're
 278          * just using the normal key deletion path:
 279          */
 280         if (bkey_deleted(&n->k))
 281                 n->k.size = 0;
 282
 283         return bch2_trans_relock(trans) ?:
 284                 bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 285                 bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
 286 }
 287
 288 static int bch2_move_extent(struct btree_trans *trans,
 289                             struct btree_iter *iter,
 290                             struct moving_context *ctxt,
 291                             struct move_bucket_in_flight *bucket_in_flight,
 292                             struct bch_io_opts io_opts,
 293                             enum btree_id btree_id,
 294                             struct bkey_s_c k,
 295                             struct data_update_opts data_opts)
 296 {
 297         struct bch_fs *c = trans->c;
 298         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 299         struct moving_io *io;
 300         const union bch_extent_entry *entry;
 301         struct extent_ptr_decoded p;
 302         unsigned sectors = k.k->size, pages;
 303         int ret = -ENOMEM;
 304
 305         trace_move_extent2(c, k);
 306
 307         bch2_data_update_opts_normalize(k, &data_opts);
 308
 309         if (!data_opts.rewrite_ptrs &&
 310             !data_opts.extra_replicas) {
 311                 if (data_opts.kill_ptrs)
 312                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 313                 return 0;
 314         }
 315
 316         /*
 317          * Before memory allocations & taking nocow locks in
 318          * bch2_data_update_init():
 319          */
 320         bch2_trans_unlock(trans);
 321
 322         /* write path might have to decompress data: */
 323         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 324                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 325
 326         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 327         io = kzalloc(sizeof(struct moving_io) +
 328                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 329         if (!io)
 330                 goto err;
 331
 332         INIT_LIST_HEAD(&io->io_list);
 333         io->write.ctxt          = ctxt;
 334         io->read_sectors        = k.k->size;
 335         io->write_sectors       = k.k->size;
 336
 337         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 338         bio_set_prio(&io->write.op.wbio.bio,
 339                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 340
 341         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 342                                  GFP_KERNEL))
 343                 goto err_free;
 344
 345         io->rbio.c              = c;
 346         io->rbio.opts           = io_opts;
 347         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 348         io->rbio.bio.bi_vcnt = pages;
 349         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 350         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 351
 352         io->rbio.bio.bi_opf             = REQ_OP_READ;
 353         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 354         io->rbio.bio.bi_end_io          = move_read_endio;
 355
 356         ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
 357                                     io_opts, data_opts, btree_id, k);
 358         if (ret && ret != -BCH_ERR_unwritten_extent_update)
 359                 goto err_free_pages;
 360
 361         if (ret == -BCH_ERR_unwritten_extent_update) {
 362                 bch2_update_unwritten_extent(trans, &io->write);
 363                 move_free(io);
 364                 return 0;
 365         }
 366
 367         BUG_ON(ret);
 368
 369         io->write.ctxt = ctxt;
 370         io->write.op.end_io = move_write_done;
 371
 372         if (ctxt->stats) {
 373                 atomic64_inc(&ctxt->stats->keys_moved);
 374                 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 375         }
 376
 377         if (bucket_in_flight) {
 378                 io->b = bucket_in_flight;
 379                 atomic_inc(&io->b->count);
 380         }
 381
 382         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 383         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 384         trace_move_extent_read2(c, k);
 385
 386         mutex_lock(&ctxt->lock);
 387         atomic_add(io->read_sectors, &ctxt->read_sectors);
 388         atomic_inc(&ctxt->read_ios);
 389
 390         list_add_tail(&io->read_list, &ctxt->reads);
 391         list_add_tail(&io->io_list, &ctxt->ios);
 392         mutex_unlock(&ctxt->lock);
 393
 394         /*
 395          * dropped by move_read_endio() - guards against use after free of
 396          * ctxt when doing wakeup
 397          */
 398         closure_get(&ctxt->cl);
 399         bch2_read_extent(trans, &io->rbio,
 400                          bkey_start_pos(k.k),
 401                          btree_id, k, 0,
 402                          BCH_READ_NODECODE|
 403                          BCH_READ_LAST_FRAGMENT);
 404         return 0;
 405 err_free_pages:
 406         bio_free_pages(&io->write.op.wbio.bio);
 407 err_free:
 408         kfree(io);
 409 err:
 410         this_cpu_inc(c->counters[BCH_COUNTER_move_extent_alloc_mem_fail]);
 411         trace_move_extent_alloc_mem_fail2(c, k);
 412         return ret;
 413 }
 414
 415 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 416                         struct bch_inode_unpacked *inode)
 417 {
 418         struct btree_iter iter;
 419         struct bkey_s_c k;
 420         int ret;
 421
 422         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 423                              BTREE_ITER_ALL_SNAPSHOTS);
 424         k = bch2_btree_iter_peek(&iter);
 425         ret = bkey_err(k);
 426         if (ret)
 427                 goto err;
 428
 429         if (!k.k || !bkey_eq(k.k->p, pos)) {
 430                 ret = -BCH_ERR_ENOENT_inode;
 431                 goto err;
 432         }
 433
 434         ret = bkey_is_inode(k.k) ? 0 : -EIO;
 435         if (ret)
 436                 goto err;
 437
 438         ret = bch2_inode_unpack(k, inode);
 439         if (ret)
 440                 goto err;
 441 err:
 442         bch2_trans_iter_exit(trans, &iter);
 443         return ret;
 444 }
 445
 446 static int move_ratelimit(struct btree_trans *trans,
 447                           struct moving_context *ctxt)
 448 {
 449         struct bch_fs *c = trans->c;
 450         u64 delay;
 451
 452         if (ctxt->wait_on_copygc) {
 453                 bch2_trans_unlock(trans);
 454                 wait_event_killable(c->copygc_running_wq,
 455                                     !c->copygc_running ||
 456                                     kthread_should_stop());
 457         }
 458
 459         do {
 460                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 461
 462                 if (delay) {
 463                         bch2_trans_unlock(trans);
 464                         set_current_state(TASK_INTERRUPTIBLE);
 465                 }
 466
 467                 if ((current->flags & PF_KTHREAD) && kthread_should_stop()) {
 468                         __set_current_state(TASK_RUNNING);
 469                         return 1;
 470                 }
 471
 472                 if (delay)
 473                         schedule_timeout(delay);
 474
 475                 if (unlikely(freezing(current))) {
 476                         move_ctxt_wait_event(ctxt, trans, list_empty(&ctxt->reads));
 477                         try_to_freeze();
 478                 }
 479         } while (delay);
 480
 481         /*
 482          * XXX: these limits really ought to be per device, SSDs and hard drives
 483          * will want different limits
 484          */
 485         move_ctxt_wait_event(ctxt, trans,
 486                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 487                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 488                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 489                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 490
 491         return 0;
 492 }
 493
 494 static int move_get_io_opts(struct btree_trans *trans,
 495                             struct bch_io_opts *io_opts,
 496                             struct bkey_s_c k, u64 *cur_inum)
 497 {
 498         struct bch_inode_unpacked inode;
 499         int ret;
 500
 501         if (*cur_inum == k.k->p.inode)
 502                 return 0;
 503
 504         ret = lookup_inode(trans,
 505                            SPOS(0, k.k->p.inode, k.k->p.snapshot),
 506                            &inode);
 507         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 508                 return ret;
 509
 510         if (!ret)
 511                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 512         else
 513                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 514         *cur_inum = k.k->p.inode;
 515         return 0;
 516 }
 517
 518 static int __bch2_move_data(struct moving_context *ctxt,
 519                             struct bpos start,
 520                             struct bpos end,
 521                             move_pred_fn pred, void *arg,
 522                             enum btree_id btree_id)
 523 {
 524         struct bch_fs *c = ctxt->c;
 525         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 526         struct bkey_buf sk;
 527         struct btree_trans trans;
 528         struct btree_iter iter;
 529         struct bkey_s_c k;
 530         struct data_update_opts data_opts;
 531         u64 cur_inum = U64_MAX;
 532         int ret = 0, ret2;
 533
 534         bch2_bkey_buf_init(&sk);
 535         bch2_trans_init(&trans, c, 0, 0);
 536
 537         if (ctxt->stats) {
 538                 ctxt->stats->data_type  = BCH_DATA_user;
 539                 ctxt->stats->btree_id   = btree_id;
 540                 ctxt->stats->pos        = start;
 541         }
 542
 543         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 544                              BTREE_ITER_PREFETCH|
 545                              BTREE_ITER_ALL_SNAPSHOTS);
 546
 547         if (ctxt->rate)
 548                 bch2_ratelimit_reset(ctxt->rate);
 549
 550         while (!move_ratelimit(&trans, ctxt)) {
 551                 bch2_trans_begin(&trans);
 552
 553                 k = bch2_btree_iter_peek(&iter);
 554                 if (!k.k)
 555                         break;
 556
 557                 ret = bkey_err(k);
 558                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 559                         continue;
 560                 if (ret)
 561                         break;
 562
 563                 if (bkey_ge(bkey_start_pos(k.k), end))
 564                         break;
 565
 566                 if (ctxt->stats)
 567                         ctxt->stats->pos = iter.pos;
 568
 569                 if (!bkey_extent_is_direct_data(k.k))
 570                         goto next_nondata;
 571
 572                 ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 573                 if (ret)
 574                         continue;
 575
 576                 memset(&data_opts, 0, sizeof(data_opts));
 577                 if (!pred(c, arg, k, &io_opts, &data_opts))
 578                         goto next;
 579
 580                 /*
 581                  * The iterator gets unlocked by __bch2_read_extent - need to
 582                  * save a copy of @k elsewhere:
 583                  */
 584                 bch2_bkey_buf_reassemble(&sk, c, k);
 585                 k = bkey_i_to_s_c(sk.k);
 586
 587                 ret2 = bch2_move_extent(&trans, &iter, ctxt, NULL,
 588                                         io_opts, btree_id, k, data_opts);
 589                 if (ret2) {
 590                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 591                                 continue;
 592
 593                         if (ret2 == -ENOMEM) {
 594                                 /* memory allocation failure, wait for some IO to finish */
 595                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 596                                 continue;
 597                         }
 598
 599                         /* XXX signal failure */
 600                         goto next;
 601                 }
 602
 603                 if (ctxt->rate)
 604                         bch2_ratelimit_increment(ctxt->rate, k.k->size);
 605 next:
 606                 if (ctxt->stats)
 607                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 608 next_nondata:
 609                 bch2_btree_iter_advance(&iter);
 610         }
 611
 612         bch2_trans_iter_exit(&trans, &iter);
 613         bch2_trans_exit(&trans);
 614         bch2_bkey_buf_exit(&sk, c);
 615
 616         return ret;
 617 }
 618
 619 int bch2_move_data(struct bch_fs *c,
 620                    enum btree_id start_btree_id, struct bpos start_pos,
 621                    enum btree_id end_btree_id,   struct bpos end_pos,
 622                    struct bch_ratelimit *rate,
 623                    struct bch_move_stats *stats,
 624                    struct write_point_specifier wp,
 625                    bool wait_on_copygc,
 626                    move_pred_fn pred, void *arg)
 627 {
 628         struct moving_context ctxt;
 629         enum btree_id id;
 630         int ret;
 631
 632         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 633
 634         for (id = start_btree_id;
 635              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 636              id++) {
 637                 stats->btree_id = id;
 638
 639                 if (id != BTREE_ID_extents &&
 640                     id != BTREE_ID_reflink)
 641                         continue;
 642
 643                 ret = __bch2_move_data(&ctxt,
 644                                        id == start_btree_id ? start_pos : POS_MIN,
 645                                        id == end_btree_id   ? end_pos   : POS_MAX,
 646                                        pred, arg, id);
 647                 if (ret)
 648                         break;
 649         }
 650
 651         bch2_moving_ctxt_exit(&ctxt);
 652
 653         return ret;
 654 }
 655
 656 int __bch2_evacuate_bucket(struct btree_trans *trans,
 657                            struct moving_context *ctxt,
 658                            struct move_bucket_in_flight *bucket_in_flight,
 659                            struct bpos bucket, int gen,
 660                            struct data_update_opts _data_opts)
 661 {
 662         struct bch_fs *c = ctxt->c;
 663         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 664         struct btree_iter iter;
 665         struct bkey_buf sk;
 666         struct bch_backpointer bp;
 667         struct bch_alloc_v4 a_convert;
 668         const struct bch_alloc_v4 *a;
 669         struct bkey_s_c k;
 670         struct data_update_opts data_opts;
 671         unsigned dirty_sectors, bucket_size;
 672         u64 fragmentation;
 673         u64 cur_inum = U64_MAX;
 674         struct bpos bp_pos = POS_MIN;
 675         int ret = 0;
 676
 677         trace_bucket_evacuate(c, &bucket);
 678
 679         bch2_bkey_buf_init(&sk);
 680
 681         /*
 682          * We're not run in a context that handles transaction restarts:
 683          */
 684         bch2_trans_begin(trans);
 685
 686         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 687                              bucket, BTREE_ITER_CACHED);
 688         ret = lockrestart_do(trans,
 689                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 690         bch2_trans_iter_exit(trans, &iter);
 691
 692         if (ret) {
 693                 bch_err(c, "%s: error looking up alloc key: %s", __func__, bch2_err_str(ret));
 694                 goto err;
 695         }
 696
 697         a = bch2_alloc_to_v4(k, &a_convert);
 698         dirty_sectors = a->dirty_sectors;
 699         bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 700         fragmentation = a->fragmentation_lru;
 701
 702         ret = bch2_btree_write_buffer_flush(trans);
 703         if (ret) {
 704                 bch_err(c, "%s: error flushing btree write buffer: %s", __func__, bch2_err_str(ret));
 705                 goto err;
 706         }
 707
 708         while (!(ret = move_ratelimit(trans, ctxt))) {
 709                 bch2_trans_begin(trans);
 710
 711                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 712                                                 &bp_pos, &bp,
 713                                                 BTREE_ITER_CACHED);
 714                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 715                         continue;
 716                 if (ret)
 717                         goto err;
 718                 if (bkey_eq(bp_pos, POS_MAX))
 719                         break;
 720
 721                 if (!bp.level) {
 722                         const struct bch_extent_ptr *ptr;
 723                         struct bkey_s_c k;
 724                         unsigned i = 0;
 725
 726                         k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
 727                         ret = bkey_err(k);
 728                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 729                                 continue;
 730                         if (ret)
 731                                 goto err;
 732                         if (!k.k)
 733                                 goto next;
 734
 735                         bch2_bkey_buf_reassemble(&sk, c, k);
 736                         k = bkey_i_to_s_c(sk.k);
 737
 738                         ret = move_get_io_opts(trans, &io_opts, k, &cur_inum);
 739                         if (ret) {
 740                                 bch2_trans_iter_exit(trans, &iter);
 741                                 continue;
 742                         }
 743
 744                         data_opts = _data_opts;
 745                         data_opts.target        = io_opts.background_target;
 746                         data_opts.rewrite_ptrs = 0;
 747
 748                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 749                                 if (ptr->dev == bucket.inode) {
 750                                         data_opts.rewrite_ptrs |= 1U << i;
 751                                         if (ptr->cached) {
 752                                                 bch2_trans_iter_exit(trans, &iter);
 753                                                 goto next;
 754                                         }
 755                                 }
 756                                 i++;
 757                         }
 758
 759                         ret = bch2_move_extent(trans, &iter, ctxt,
 760                                         bucket_in_flight,
 761                                         io_opts, bp.btree_id, k, data_opts);
 762                         bch2_trans_iter_exit(trans, &iter);
 763
 764                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 765                                 continue;
 766                         if (ret == -ENOMEM) {
 767                                 /* memory allocation failure, wait for some IO to finish */
 768                                 bch2_move_ctxt_wait_for_io(ctxt, trans);
 769                                 continue;
 770                         }
 771                         if (ret)
 772                                 goto err;
 773
 774                         if (ctxt->rate)
 775                                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 776                         if (ctxt->stats)
 777                                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 778                 } else {
 779                         struct btree *b;
 780
 781                         b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
 782                         ret = PTR_ERR_OR_ZERO(b);
 783                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 784                                 continue;
 785                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 786                                 continue;
 787                         if (ret)
 788                                 goto err;
 789                         if (!b)
 790                                 goto next;
 791
 792                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 793                         bch2_trans_iter_exit(trans, &iter);
 794
 795                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 796                                 continue;
 797                         if (ret)
 798                                 goto err;
 799
 800                         if (ctxt->rate)
 801                                 bch2_ratelimit_increment(ctxt->rate,
 802                                                          c->opts.btree_node_size >> 9);
 803                         if (ctxt->stats) {
 804                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 805                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 806                         }
 807                 }
 808 next:
 809                 bp_pos = bpos_nosnap_successor(bp_pos);
 810         }
 811
 812         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 813 err:
 814         bch2_bkey_buf_exit(&sk, c);
 815         return ret;
 816 }
 817
 818 int bch2_evacuate_bucket(struct bch_fs *c,
 819                          struct bpos bucket, int gen,
 820                          struct data_update_opts data_opts,
 821                          struct bch_ratelimit *rate,
 822                          struct bch_move_stats *stats,
 823                          struct write_point_specifier wp,
 824                          bool wait_on_copygc)
 825 {
 826         struct btree_trans trans;
 827         struct moving_context ctxt;
 828         int ret;
 829
 830         bch2_trans_init(&trans, c, 0, 0);
 831         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 832         ret = __bch2_evacuate_bucket(&trans, &ctxt, NULL, bucket, gen, data_opts);
 833         bch2_moving_ctxt_exit(&ctxt);
 834         bch2_trans_exit(&trans);
 835
 836         return ret;
 837 }
 838
 839 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 840                                 struct btree *, struct bch_io_opts *,
 841                                 struct data_update_opts *);
 842
 843 static int bch2_move_btree(struct bch_fs *c,
 844                            enum btree_id start_btree_id, struct bpos start_pos,
 845                            enum btree_id end_btree_id,   struct bpos end_pos,
 846                            move_btree_pred pred, void *arg,
 847                            struct bch_move_stats *stats)
 848 {
 849         bool kthread = (current->flags & PF_KTHREAD) != 0;
 850         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 851         struct btree_trans trans;
 852         struct btree_iter iter;
 853         struct btree *b;
 854         enum btree_id id;
 855         struct data_update_opts data_opts;
 856         int ret = 0;
 857
 858         bch2_trans_init(&trans, c, 0, 0);
 859         progress_list_add(c, stats);
 860
 861         stats->data_type = BCH_DATA_btree;
 862
 863         for (id = start_btree_id;
 864              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 865              id++) {
 866                 stats->btree_id = id;
 867
 868                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 869                                           BTREE_ITER_PREFETCH);
 870 retry:
 871                 ret = 0;
 872                 while (bch2_trans_begin(&trans),
 873                        (b = bch2_btree_iter_peek_node(&iter)) &&
 874                        !(ret = PTR_ERR_OR_ZERO(b))) {
 875                         if (kthread && kthread_should_stop())
 876                                 break;
 877
 878                         if ((cmp_int(id, end_btree_id) ?:
 879                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 880                                 break;
 881
 882                         stats->pos = iter.pos;
 883
 884                         if (!pred(c, arg, b, &io_opts, &data_opts))
 885                                 goto next;
 886
 887                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 888                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 889                                 continue;
 890                         if (ret)
 891                                 break;
 892 next:
 893                         bch2_btree_iter_next_node(&iter);
 894                 }
 895                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 896                         goto retry;
 897
 898                 bch2_trans_iter_exit(&trans, &iter);
 899
 900                 if (kthread && kthread_should_stop())
 901                         break;
 902         }
 903
 904         bch2_trans_exit(&trans);
 905
 906         if (ret)
 907                 bch_err(c, "error in %s(): %s", __func__, bch2_err_str(ret));
 908
 909         bch2_btree_interior_updates_flush(c);
 910
 911         progress_list_del(c, stats);
 912         return ret;
 913 }
 914
 915 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 916                              struct bkey_s_c k,
 917                              struct bch_io_opts *io_opts,
 918                              struct data_update_opts *data_opts)
 919 {
 920         unsigned nr_good = bch2_bkey_durability(c, k);
 921         unsigned replicas = bkey_is_btree_ptr(k.k)
 922                 ? c->opts.metadata_replicas
 923                 : io_opts->data_replicas;
 924
 925         if (!nr_good || nr_good >= replicas)
 926                 return false;
 927
 928         data_opts->target               = 0;
 929         data_opts->extra_replicas       = replicas - nr_good;
 930         data_opts->btree_insert_flags   = 0;
 931         return true;
 932 }
 933
 934 static bool migrate_pred(struct bch_fs *c, void *arg,
 935                          struct bkey_s_c k,
 936                          struct bch_io_opts *io_opts,
 937                          struct data_update_opts *data_opts)
 938 {
 939         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 940         const struct bch_extent_ptr *ptr;
 941         struct bch_ioctl_data *op = arg;
 942         unsigned i = 0;
 943
 944         data_opts->rewrite_ptrs         = 0;
 945         data_opts->target               = 0;
 946         data_opts->extra_replicas       = 0;
 947         data_opts->btree_insert_flags   = 0;
 948
 949         bkey_for_each_ptr(ptrs, ptr) {
 950                 if (ptr->dev == op->migrate.dev)
 951                         data_opts->rewrite_ptrs |= 1U << i;
 952                 i++;
 953         }
 954
 955         return data_opts->rewrite_ptrs != 0;
 956 }
 957
 958 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 959                                    struct btree *b,
 960                                    struct bch_io_opts *io_opts,
 961                                    struct data_update_opts *data_opts)
 962 {
 963         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 964 }
 965
 966 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 967                                struct btree *b,
 968                                struct bch_io_opts *io_opts,
 969                                struct data_update_opts *data_opts)
 970 {
 971         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 972 }
 973
 974 static bool bformat_needs_redo(struct bkey_format *f)
 975 {
 976         unsigned i;
 977
 978         for (i = 0; i < f->nr_fields; i++) {
 979                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 980                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 981                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 982
 983                 if (f->bits_per_field[i] > unpacked_bits)
 984                         return true;
 985
 986                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 987                         return true;
 988
 989                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 990                      unpacked_mask) <
 991                     field_offset)
 992                         return true;
 993         }
 994
 995         return false;
 996 }
 997
 998 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 999                                    struct btree *b,
1000                                    struct bch_io_opts *io_opts,
1001                                    struct data_update_opts *data_opts)
1002 {
1003         if (b->version_ondisk != c->sb.version ||
1004             btree_node_need_rewrite(b) ||
1005             bformat_needs_redo(&b->format)) {
1006                 data_opts->target               = 0;
1007                 data_opts->extra_replicas       = 0;
1008                 data_opts->btree_insert_flags   = 0;
1009                 return true;
1010         }
1011
1012         return false;
1013 }
1014
1015 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1016 {
1017         int ret;
1018
1019         ret = bch2_move_btree(c,
1020                               0,                POS_MIN,
1021                               BTREE_ID_NR,      SPOS_MAX,
1022                               rewrite_old_nodes_pred, c, stats);
1023         if (!ret) {
1024                 mutex_lock(&c->sb_lock);
1025                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1026                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1027                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1028                 bch2_write_super(c);
1029                 mutex_unlock(&c->sb_lock);
1030         }
1031
1032         return ret;
1033 }
1034
1035 int bch2_data_job(struct bch_fs *c,
1036                   struct bch_move_stats *stats,
1037                   struct bch_ioctl_data op)
1038 {
1039         int ret = 0;
1040
1041         switch (op.op) {
1042         case BCH_DATA_OP_REREPLICATE:
1043                 bch2_move_stats_init(stats, "rereplicate");
1044                 stats->data_type = BCH_DATA_journal;
1045                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1046
1047                 ret = bch2_move_btree(c,
1048                                       op.start_btree,   op.start_pos,
1049                                       op.end_btree,     op.end_pos,
1050                                       rereplicate_btree_pred, c, stats) ?: ret;
1051                 ret = bch2_replicas_gc2(c) ?: ret;
1052
1053                 ret = bch2_move_data(c,
1054                                      op.start_btree,    op.start_pos,
1055                                      op.end_btree,      op.end_pos,
1056                                      NULL,
1057                                      stats,
1058                                      writepoint_hashed((unsigned long) current),
1059                                      true,
1060                                      rereplicate_pred, c) ?: ret;
1061                 ret = bch2_replicas_gc2(c) ?: ret;
1062                 break;
1063         case BCH_DATA_OP_MIGRATE:
1064                 if (op.migrate.dev >= c->sb.nr_devices)
1065                         return -EINVAL;
1066
1067                 bch2_move_stats_init(stats, "migrate");
1068                 stats->data_type = BCH_DATA_journal;
1069                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1070
1071                 ret = bch2_move_btree(c,
1072                                       op.start_btree,   op.start_pos,
1073                                       op.end_btree,     op.end_pos,
1074                                       migrate_btree_pred, &op, stats) ?: ret;
1075                 ret = bch2_replicas_gc2(c) ?: ret;
1076
1077                 ret = bch2_move_data(c,
1078                                      op.start_btree,    op.start_pos,
1079                                      op.end_btree,      op.end_pos,
1080                                      NULL,
1081                                      stats,
1082                                      writepoint_hashed((unsigned long) current),
1083                                      true,
1084                                      migrate_pred, &op) ?: ret;
1085                 ret = bch2_replicas_gc2(c) ?: ret;
1086                 break;
1087         case BCH_DATA_OP_REWRITE_OLD_NODES:
1088                 bch2_move_stats_init(stats, "rewrite_old_nodes");
1089                 ret = bch2_scan_old_btree_nodes(c, stats);
1090                 break;
1091         default:
1092                 ret = -EINVAL;
1093         }
1094
1095         return ret;
1096 }
1097
1098 void bch2_data_jobs_to_text(struct printbuf *out, struct bch_fs *c)
1099 {
1100         struct bch_move_stats *stats;
1101
1102         mutex_lock(&c->data_progress_lock);
1103         list_for_each_entry(stats, &c->data_progress_list, list) {
1104                 prt_printf(out, "%s: data type %s btree_id %s position: ",
1105                        stats->name,
1106                        bch2_data_types[stats->data_type],
1107                        bch2_btree_ids[stats->btree_id]);
1108                 bch2_bpos_to_text(out, stats->pos);
1109                 prt_printf(out, "%s", "\n");
1110         }
1111         mutex_unlock(&c->data_progress_lock);
1112 }
1113
1114 static void bch2_moving_ctxt_to_text(struct printbuf *out, struct moving_context *ctxt)
1115 {
1116         struct moving_io *io;
1117
1118         prt_printf(out, "%ps:", ctxt->fn);
1119         prt_newline(out);
1120         printbuf_indent_add(out, 2);
1121
1122         prt_printf(out, "reads: %u sectors %u",
1123                    atomic_read(&ctxt->read_ios),
1124                    atomic_read(&ctxt->read_sectors));
1125         prt_newline(out);
1126
1127         prt_printf(out, "writes: %u sectors %u",
1128                    atomic_read(&ctxt->write_ios),
1129                    atomic_read(&ctxt->write_sectors));
1130         prt_newline(out);
1131
1132         printbuf_indent_add(out, 2);
1133
1134         mutex_lock(&ctxt->lock);
1135         list_for_each_entry(io, &ctxt->ios, io_list) {
1136                 bch2_write_op_to_text(out, &io->write.op);
1137         }
1138         mutex_unlock(&ctxt->lock);
1139
1140         printbuf_indent_sub(out, 4);
1141 }
1142
1143 void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
1144 {
1145         struct moving_context *ctxt;
1146
1147         mutex_lock(&c->moving_context_lock);
1148         list_for_each_entry(ctxt, &c->moving_context_list, list)
1149                 bch2_moving_ctxt_to_text(out, ctxt);
1150         mutex_unlock(&c->moving_context_lock);
1151 }
1152
1153 void bch2_fs_move_init(struct bch_fs *c)
1154 {
1155         INIT_LIST_HEAD(&c->moving_context_list);
1156         mutex_init(&c->moving_context_lock);
1157
1158         INIT_LIST_HEAD(&c->data_progress_list);
1159         mutex_init(&c->data_progress_lock);
1160 }