fs/bcachefs/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_io.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "btree_write_buffer.h"
  13 #include "compress.h"
  14 #include "disk_groups.h"
  15 #include "ec.h"
  16 #include "errcode.h"
  17 #include "error.h"
  18 #include "inode.h"
  19 #include "io_read.h"
  20 #include "io_write.h"
  21 #include "journal_reclaim.h"
  22 #include "keylist.h"
  23 #include "move.h"
  24 #include "replicas.h"
  25 #include "snapshot.h"
  26 #include "super-io.h"
  27 #include "trace.h"
  28
  29 #include <linux/ioprio.h>
  30 #include <linux/kthread.h>
  31
  32 const char * const bch2_data_ops_strs[] = {
  33 #define x(t, n, ...) [n] = #t,
  34         BCH_DATA_OPS()
  35 #undef x
  36         NULL
  37 };
  38
  39 static void trace_move_extent2(struct bch_fs *c, struct bkey_s_c k,
  40                                struct bch_io_opts *io_opts,
  41                                struct data_update_opts *data_opts)
  42 {
  43         if (trace_move_extent_enabled()) {
  44                 struct printbuf buf = PRINTBUF;
  45
  46                 bch2_bkey_val_to_text(&buf, c, k);
  47                 prt_newline(&buf);
  48                 bch2_data_update_opts_to_text(&buf, c, io_opts, data_opts);
  49                 trace_move_extent(c, buf.buf);
  50                 printbuf_exit(&buf);
  51         }
  52 }
  53
  54 static void trace_move_extent_read2(struct bch_fs *c, struct bkey_s_c k)
  55 {
  56         if (trace_move_extent_read_enabled()) {
  57                 struct printbuf buf = PRINTBUF;
  58
  59                 bch2_bkey_val_to_text(&buf, c, k);
  60                 trace_move_extent_read(c, buf.buf);
  61                 printbuf_exit(&buf);
  62         }
  63 }
  64
  65 struct moving_io {
  66         struct list_head                read_list;
  67         struct list_head                io_list;
  68         struct move_bucket_in_flight    *b;
  69         struct closure                  cl;
  70         bool                            read_completed;
  71
  72         unsigned                        read_sectors;
  73         unsigned                        write_sectors;
  74
  75         struct bch_read_bio             rbio;
  76
  77         struct data_update              write;
  78         /* Must be last since it is variable size */
  79         struct bio_vec                  bi_inline_vecs[];
  80 };
  81
  82 static void move_free(struct moving_io *io)
  83 {
  84         struct moving_context *ctxt = io->write.ctxt;
  85
  86         if (io->b)
  87                 atomic_dec(&io->b->count);
  88
  89         bch2_data_update_exit(&io->write);
  90
  91         mutex_lock(&ctxt->lock);
  92         list_del(&io->io_list);
  93         wake_up(&ctxt->wait);
  94         mutex_unlock(&ctxt->lock);
  95
  96         kfree(io);
  97 }
  98
  99 static void move_write_done(struct bch_write_op *op)
 100 {
 101         struct moving_io *io = container_of(op, struct moving_io, write.op);
 102         struct moving_context *ctxt = io->write.ctxt;
 103
 104         if (io->write.op.error)
 105                 ctxt->write_error = true;
 106
 107         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 108         atomic_dec(&io->write.ctxt->write_ios);
 109         move_free(io);
 110         closure_put(&ctxt->cl);
 111 }
 112
 113 static void move_write(struct moving_io *io)
 114 {
 115         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 116                 move_free(io);
 117                 return;
 118         }
 119
 120         if (trace_move_extent_write_enabled()) {
 121                 struct bch_fs *c = io->write.op.c;
 122                 struct printbuf buf = PRINTBUF;
 123
 124                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(io->write.k.k));
 125                 trace_move_extent_write(c, buf.buf);
 126                 printbuf_exit(&buf);
 127         }
 128
 129         closure_get(&io->write.ctxt->cl);
 130         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 131         atomic_inc(&io->write.ctxt->write_ios);
 132
 133         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
 134 }
 135
 136 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
 137 {
 138         struct moving_io *io =
 139                 list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
 140
 141         return io && io->read_completed ? io : NULL;
 142 }
 143
 144 static void move_read_endio(struct bio *bio)
 145 {
 146         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 147         struct moving_context *ctxt = io->write.ctxt;
 148
 149         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 150         atomic_dec(&ctxt->read_ios);
 151         io->read_completed = true;
 152
 153         wake_up(&ctxt->wait);
 154         closure_put(&ctxt->cl);
 155 }
 156
 157 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt)
 158 {
 159         struct moving_io *io;
 160
 161         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 162                 bch2_trans_unlock_long(ctxt->trans);
 163                 list_del(&io->read_list);
 164                 move_write(io);
 165         }
 166 }
 167
 168 void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 169 {
 170         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 171
 172         move_ctxt_wait_event(ctxt,
 173                 !atomic_read(&ctxt->write_sectors) ||
 174                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 175 }
 176
 177 void bch2_moving_ctxt_flush_all(struct moving_context *ctxt)
 178 {
 179         move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 180         bch2_trans_unlock_long(ctxt->trans);
 181         closure_sync(&ctxt->cl);
 182 }
 183
 184 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 185 {
 186         struct bch_fs *c = ctxt->trans->c;
 187
 188         bch2_moving_ctxt_flush_all(ctxt);
 189
 190         EBUG_ON(atomic_read(&ctxt->write_sectors));
 191         EBUG_ON(atomic_read(&ctxt->write_ios));
 192         EBUG_ON(atomic_read(&ctxt->read_sectors));
 193         EBUG_ON(atomic_read(&ctxt->read_ios));
 194
 195         mutex_lock(&c->moving_context_lock);
 196         list_del(&ctxt->list);
 197         mutex_unlock(&c->moving_context_lock);
 198
 199         bch2_trans_put(ctxt->trans);
 200         memset(ctxt, 0, sizeof(*ctxt));
 201 }
 202
 203 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 204                            struct bch_fs *c,
 205                            struct bch_ratelimit *rate,
 206                            struct bch_move_stats *stats,
 207                            struct write_point_specifier wp,
 208                            bool wait_on_copygc)
 209 {
 210         memset(ctxt, 0, sizeof(*ctxt));
 211
 212         ctxt->trans     = bch2_trans_get(c);
 213         ctxt->fn        = (void *) _RET_IP_;
 214         ctxt->rate      = rate;
 215         ctxt->stats     = stats;
 216         ctxt->wp        = wp;
 217         ctxt->wait_on_copygc = wait_on_copygc;
 218
 219         closure_init_stack(&ctxt->cl);
 220
 221         mutex_init(&ctxt->lock);
 222         INIT_LIST_HEAD(&ctxt->reads);
 223         INIT_LIST_HEAD(&ctxt->ios);
 224         init_waitqueue_head(&ctxt->wait);
 225
 226         mutex_lock(&c->moving_context_lock);
 227         list_add(&ctxt->list, &c->moving_context_list);
 228         mutex_unlock(&c->moving_context_lock);
 229 }
 230
 231 void bch2_move_stats_exit(struct bch_move_stats *stats, struct bch_fs *c)
 232 {
 233         trace_move_data(c, stats);
 234 }
 235
 236 void bch2_move_stats_init(struct bch_move_stats *stats, const char *name)
 237 {
 238         memset(stats, 0, sizeof(*stats));
 239         stats->data_type = BCH_DATA_user;
 240         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 241 }
 242
 243 int bch2_move_extent(struct moving_context *ctxt,
 244                      struct move_bucket_in_flight *bucket_in_flight,
 245                      struct btree_iter *iter,
 246                      struct bkey_s_c k,
 247                      struct bch_io_opts io_opts,
 248                      struct data_update_opts data_opts)
 249 {
 250         struct btree_trans *trans = ctxt->trans;
 251         struct bch_fs *c = trans->c;
 252         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 253         struct moving_io *io;
 254         const union bch_extent_entry *entry;
 255         struct extent_ptr_decoded p;
 256         unsigned sectors = k.k->size, pages;
 257         int ret = -ENOMEM;
 258
 259         trace_move_extent2(c, k, &io_opts, &data_opts);
 260
 261         if (ctxt->stats)
 262                 ctxt->stats->pos = BBPOS(iter->btree_id, iter->pos);
 263
 264         bch2_data_update_opts_normalize(k, &data_opts);
 265
 266         if (!data_opts.rewrite_ptrs &&
 267             !data_opts.extra_replicas) {
 268                 if (data_opts.kill_ptrs)
 269                         return bch2_extent_drop_ptrs(trans, iter, k, &io_opts, &data_opts);
 270                 return 0;
 271         }
 272
 273         /*
 274          * Before memory allocations & taking nocow locks in
 275          * bch2_data_update_init():
 276          */
 277         bch2_trans_unlock(trans);
 278
 279         /* write path might have to decompress data: */
 280         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 281                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 282
 283         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 284         io = kzalloc(sizeof(struct moving_io) +
 285                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 286         if (!io)
 287                 goto err;
 288
 289         INIT_LIST_HEAD(&io->io_list);
 290         io->write.ctxt          = ctxt;
 291         io->read_sectors        = k.k->size;
 292         io->write_sectors       = k.k->size;
 293
 294         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 295         bio_set_prio(&io->write.op.wbio.bio,
 296                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 297
 298         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 299                                  GFP_KERNEL))
 300                 goto err_free;
 301
 302         io->rbio.c              = c;
 303         io->rbio.opts           = io_opts;
 304         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 305         io->rbio.bio.bi_vcnt = pages;
 306         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 307         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 308
 309         io->rbio.bio.bi_opf             = REQ_OP_READ;
 310         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 311         io->rbio.bio.bi_end_io          = move_read_endio;
 312
 313         ret = bch2_data_update_init(trans, iter, ctxt, &io->write, ctxt->wp,
 314                                     io_opts, data_opts, iter->btree_id, k);
 315         if (ret)
 316                 goto err_free_pages;
 317
 318         io->write.op.end_io = move_write_done;
 319
 320         if (ctxt->rate)
 321                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 322
 323         if (ctxt->stats) {
 324                 atomic64_inc(&ctxt->stats->keys_moved);
 325                 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 326         }
 327
 328         if (bucket_in_flight) {
 329                 io->b = bucket_in_flight;
 330                 atomic_inc(&io->b->count);
 331         }
 332
 333         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 334         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 335         trace_move_extent_read2(c, k);
 336
 337         mutex_lock(&ctxt->lock);
 338         atomic_add(io->read_sectors, &ctxt->read_sectors);
 339         atomic_inc(&ctxt->read_ios);
 340
 341         list_add_tail(&io->read_list, &ctxt->reads);
 342         list_add_tail(&io->io_list, &ctxt->ios);
 343         mutex_unlock(&ctxt->lock);
 344
 345         /*
 346          * dropped by move_read_endio() - guards against use after free of
 347          * ctxt when doing wakeup
 348          */
 349         closure_get(&ctxt->cl);
 350         bch2_read_extent(trans, &io->rbio,
 351                          bkey_start_pos(k.k),
 352                          iter->btree_id, k, 0,
 353                          BCH_READ_NODECODE|
 354                          BCH_READ_LAST_FRAGMENT);
 355         return 0;
 356 err_free_pages:
 357         bio_free_pages(&io->write.op.wbio.bio);
 358 err_free:
 359         kfree(io);
 360 err:
 361         if (ret == -BCH_ERR_data_update_done)
 362                 return 0;
 363
 364         if (bch2_err_matches(ret, EROFS) ||
 365             bch2_err_matches(ret, BCH_ERR_transaction_restart))
 366                 return ret;
 367
 368         count_event(c, move_extent_start_fail);
 369
 370         if (trace_move_extent_start_fail_enabled()) {
 371                 struct printbuf buf = PRINTBUF;
 372
 373                 bch2_bkey_val_to_text(&buf, c, k);
 374                 prt_str(&buf, ": ");
 375                 prt_str(&buf, bch2_err_str(ret));
 376                 trace_move_extent_start_fail(c, buf.buf);
 377                 printbuf_exit(&buf);
 378         }
 379         return ret;
 380 }
 381
 382 struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
 383                           struct per_snapshot_io_opts *io_opts,
 384                           struct bkey_s_c extent_k)
 385 {
 386         struct bch_fs *c = trans->c;
 387         u32 restart_count = trans->restart_count;
 388         int ret = 0;
 389
 390         if (io_opts->cur_inum != extent_k.k->p.inode) {
 391                 io_opts->d.nr = 0;
 392
 393                 ret = for_each_btree_key(trans, iter, BTREE_ID_inodes, POS(0, extent_k.k->p.inode),
 394                                          BTREE_ITER_all_snapshots, k, ({
 395                         if (k.k->p.offset != extent_k.k->p.inode)
 396                                 break;
 397
 398                         if (!bkey_is_inode(k.k))
 399                                 continue;
 400
 401                         struct bch_inode_unpacked inode;
 402                         BUG_ON(bch2_inode_unpack(k, &inode));
 403
 404                         struct snapshot_io_opts_entry e = { .snapshot = k.k->p.snapshot };
 405                         bch2_inode_opts_get(&e.io_opts, trans->c, &inode);
 406
 407                         darray_push(&io_opts->d, e);
 408                 }));
 409                 io_opts->cur_inum = extent_k.k->p.inode;
 410         }
 411
 412         ret = ret ?: trans_was_restarted(trans, restart_count);
 413         if (ret)
 414                 return ERR_PTR(ret);
 415
 416         if (extent_k.k->p.snapshot)
 417                 darray_for_each(io_opts->d, i)
 418                         if (bch2_snapshot_is_ancestor(c, extent_k.k->p.snapshot, i->snapshot))
 419                                 return &i->io_opts;
 420
 421         return &io_opts->fs_io_opts;
 422 }
 423
 424 int bch2_move_get_io_opts_one(struct btree_trans *trans,
 425                               struct bch_io_opts *io_opts,
 426                               struct bkey_s_c extent_k)
 427 {
 428         struct btree_iter iter;
 429         struct bkey_s_c k;
 430         int ret;
 431
 432         /* reflink btree? */
 433         if (!extent_k.k->p.inode) {
 434                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 435                 return 0;
 436         }
 437
 438         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
 439                                SPOS(0, extent_k.k->p.inode, extent_k.k->p.snapshot),
 440                                BTREE_ITER_cached);
 441         ret = bkey_err(k);
 442         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 443                 return ret;
 444
 445         if (!ret && bkey_is_inode(k.k)) {
 446                 struct bch_inode_unpacked inode;
 447                 bch2_inode_unpack(k, &inode);
 448                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 449         } else {
 450                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 451         }
 452
 453         bch2_trans_iter_exit(trans, &iter);
 454         return 0;
 455 }
 456
 457 int bch2_move_ratelimit(struct moving_context *ctxt)
 458 {
 459         struct bch_fs *c = ctxt->trans->c;
 460         bool is_kthread = current->flags & PF_KTHREAD;
 461         u64 delay;
 462
 463         if (ctxt->wait_on_copygc && c->copygc_running) {
 464                 bch2_moving_ctxt_flush_all(ctxt);
 465                 wait_event_killable(c->copygc_running_wq,
 466                                     !c->copygc_running ||
 467                                     (is_kthread && kthread_should_stop()));
 468         }
 469
 470         do {
 471                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 472
 473                 if (is_kthread && kthread_should_stop())
 474                         return 1;
 475
 476                 if (delay)
 477                         move_ctxt_wait_event_timeout(ctxt,
 478                                         freezing(current) ||
 479                                         (is_kthread && kthread_should_stop()),
 480                                         delay);
 481
 482                 if (unlikely(freezing(current))) {
 483                         bch2_moving_ctxt_flush_all(ctxt);
 484                         try_to_freeze();
 485                 }
 486         } while (delay);
 487
 488         /*
 489          * XXX: these limits really ought to be per device, SSDs and hard drives
 490          * will want different limits
 491          */
 492         move_ctxt_wait_event(ctxt,
 493                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 494                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 495                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 496                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 497
 498         return 0;
 499 }
 500
 501 static int bch2_move_data_btree(struct moving_context *ctxt,
 502                                 struct bpos start,
 503                                 struct bpos end,
 504                                 move_pred_fn pred, void *arg,
 505                                 enum btree_id btree_id)
 506 {
 507         struct btree_trans *trans = ctxt->trans;
 508         struct bch_fs *c = trans->c;
 509         struct per_snapshot_io_opts snapshot_io_opts;
 510         struct bch_io_opts *io_opts;
 511         struct bkey_buf sk;
 512         struct btree_iter iter;
 513         struct bkey_s_c k;
 514         struct data_update_opts data_opts;
 515         int ret = 0, ret2;
 516
 517         per_snapshot_io_opts_init(&snapshot_io_opts, c);
 518         bch2_bkey_buf_init(&sk);
 519
 520         if (ctxt->stats) {
 521                 ctxt->stats->data_type  = BCH_DATA_user;
 522                 ctxt->stats->pos        = BBPOS(btree_id, start);
 523         }
 524
 525         bch2_trans_begin(trans);
 526         bch2_trans_iter_init(trans, &iter, btree_id, start,
 527                              BTREE_ITER_prefetch|
 528                              BTREE_ITER_all_snapshots);
 529
 530         if (ctxt->rate)
 531                 bch2_ratelimit_reset(ctxt->rate);
 532
 533         while (!bch2_move_ratelimit(ctxt)) {
 534                 bch2_trans_begin(trans);
 535
 536                 k = bch2_btree_iter_peek(&iter);
 537                 if (!k.k)
 538                         break;
 539
 540                 ret = bkey_err(k);
 541                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 542                         continue;
 543                 if (ret)
 544                         break;
 545
 546                 if (bkey_ge(bkey_start_pos(k.k), end))
 547                         break;
 548
 549                 if (ctxt->stats)
 550                         ctxt->stats->pos = BBPOS(iter.btree_id, iter.pos);
 551
 552                 if (!bkey_extent_is_direct_data(k.k))
 553                         goto next_nondata;
 554
 555                 io_opts = bch2_move_get_io_opts(trans, &snapshot_io_opts, k);
 556                 ret = PTR_ERR_OR_ZERO(io_opts);
 557                 if (ret)
 558                         continue;
 559
 560                 memset(&data_opts, 0, sizeof(data_opts));
 561                 if (!pred(c, arg, k, io_opts, &data_opts))
 562                         goto next;
 563
 564                 /*
 565                  * The iterator gets unlocked by __bch2_read_extent - need to
 566                  * save a copy of @k elsewhere:
 567                  */
 568                 bch2_bkey_buf_reassemble(&sk, c, k);
 569                 k = bkey_i_to_s_c(sk.k);
 570
 571                 ret2 = bch2_move_extent(ctxt, NULL, &iter, k, *io_opts, data_opts);
 572                 if (ret2) {
 573                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 574                                 continue;
 575
 576                         if (ret2 == -ENOMEM) {
 577                                 /* memory allocation failure, wait for some IO to finish */
 578                                 bch2_move_ctxt_wait_for_io(ctxt);
 579                                 continue;
 580                         }
 581
 582                         /* XXX signal failure */
 583                         goto next;
 584                 }
 585 next:
 586                 if (ctxt->stats)
 587                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 588 next_nondata:
 589                 bch2_btree_iter_advance(&iter);
 590         }
 591
 592         bch2_trans_iter_exit(trans, &iter);
 593         bch2_bkey_buf_exit(&sk, c);
 594         per_snapshot_io_opts_exit(&snapshot_io_opts);
 595
 596         return ret;
 597 }
 598
 599 int __bch2_move_data(struct moving_context *ctxt,
 600                      struct bbpos start,
 601                      struct bbpos end,
 602                      move_pred_fn pred, void *arg)
 603 {
 604         struct bch_fs *c = ctxt->trans->c;
 605         enum btree_id id;
 606         int ret = 0;
 607
 608         for (id = start.btree;
 609              id <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
 610              id++) {
 611                 ctxt->stats->pos = BBPOS(id, POS_MIN);
 612
 613                 if (!btree_type_has_ptrs(id) ||
 614                     !bch2_btree_id_root(c, id)->b)
 615                         continue;
 616
 617                 ret = bch2_move_data_btree(ctxt,
 618                                        id == start.btree ? start.pos : POS_MIN,
 619                                        id == end.btree   ? end.pos   : POS_MAX,
 620                                        pred, arg, id);
 621                 if (ret)
 622                         break;
 623         }
 624
 625         return ret;
 626 }
 627
 628 int bch2_move_data(struct bch_fs *c,
 629                    struct bbpos start,
 630                    struct bbpos end,
 631                    struct bch_ratelimit *rate,
 632                    struct bch_move_stats *stats,
 633                    struct write_point_specifier wp,
 634                    bool wait_on_copygc,
 635                    move_pred_fn pred, void *arg)
 636 {
 637
 638         struct moving_context ctxt;
 639         int ret;
 640
 641         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 642         ret = __bch2_move_data(&ctxt, start, end, pred, arg);
 643         bch2_moving_ctxt_exit(&ctxt);
 644
 645         return ret;
 646 }
 647
 648 int bch2_evacuate_bucket(struct moving_context *ctxt,
 649                            struct move_bucket_in_flight *bucket_in_flight,
 650                            struct bpos bucket, int gen,
 651                            struct data_update_opts _data_opts)
 652 {
 653         struct btree_trans *trans = ctxt->trans;
 654         struct bch_fs *c = trans->c;
 655         bool is_kthread = current->flags & PF_KTHREAD;
 656         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 657         struct btree_iter iter;
 658         struct bkey_buf sk;
 659         struct bch_backpointer bp;
 660         struct bch_alloc_v4 a_convert;
 661         const struct bch_alloc_v4 *a;
 662         struct bkey_s_c k;
 663         struct data_update_opts data_opts;
 664         unsigned dirty_sectors, bucket_size;
 665         u64 fragmentation;
 666         struct bpos bp_pos = POS_MIN;
 667         int ret = 0;
 668
 669         struct bch_dev *ca = bch2_dev_tryget(c, bucket.inode);
 670         if (!ca)
 671                 return 0;
 672
 673         trace_bucket_evacuate(c, &bucket);
 674
 675         bch2_bkey_buf_init(&sk);
 676
 677         /*
 678          * We're not run in a context that handles transaction restarts:
 679          */
 680         bch2_trans_begin(trans);
 681
 682         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 683                              bucket, BTREE_ITER_cached);
 684         ret = lockrestart_do(trans,
 685                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 686         bch2_trans_iter_exit(trans, &iter);
 687
 688         bch_err_msg(c, ret, "looking up alloc key");
 689         if (ret)
 690                 goto err;
 691
 692         a = bch2_alloc_to_v4(k, &a_convert);
 693         dirty_sectors = bch2_bucket_sectors_dirty(*a);
 694         bucket_size = ca->mi.bucket_size;
 695         fragmentation = alloc_lru_idx_fragmentation(*a, ca);
 696
 697         ret = bch2_btree_write_buffer_tryflush(trans);
 698         bch_err_msg(c, ret, "flushing btree write buffer");
 699         if (ret)
 700                 goto err;
 701
 702         while (!(ret = bch2_move_ratelimit(ctxt))) {
 703                 if (is_kthread && kthread_should_stop())
 704                         break;
 705
 706                 bch2_trans_begin(trans);
 707
 708                 ret = bch2_get_next_backpointer(trans, ca, bucket, gen,
 709                                                 &bp_pos, &bp,
 710                                                 BTREE_ITER_cached);
 711                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 712                         continue;
 713                 if (ret)
 714                         goto err;
 715                 if (bkey_eq(bp_pos, POS_MAX))
 716                         break;
 717
 718                 if (!bp.level) {
 719                         k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
 720                         ret = bkey_err(k);
 721                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 722                                 continue;
 723                         if (ret)
 724                                 goto err;
 725                         if (!k.k)
 726                                 goto next;
 727
 728                         bch2_bkey_buf_reassemble(&sk, c, k);
 729                         k = bkey_i_to_s_c(sk.k);
 730
 731                         ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
 732                         if (ret) {
 733                                 bch2_trans_iter_exit(trans, &iter);
 734                                 continue;
 735                         }
 736
 737                         data_opts = _data_opts;
 738                         data_opts.target        = io_opts.background_target;
 739                         data_opts.rewrite_ptrs = 0;
 740
 741                         unsigned i = 0;
 742                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 743                                 if (ptr->dev == bucket.inode) {
 744                                         data_opts.rewrite_ptrs |= 1U << i;
 745                                         if (ptr->cached) {
 746                                                 bch2_trans_iter_exit(trans, &iter);
 747                                                 goto next;
 748                                         }
 749                                 }
 750                                 i++;
 751                         }
 752
 753                         ret = bch2_move_extent(ctxt, bucket_in_flight,
 754                                                &iter, k, io_opts, data_opts);
 755                         bch2_trans_iter_exit(trans, &iter);
 756
 757                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 758                                 continue;
 759                         if (ret == -ENOMEM) {
 760                                 /* memory allocation failure, wait for some IO to finish */
 761                                 bch2_move_ctxt_wait_for_io(ctxt);
 762                                 continue;
 763                         }
 764                         if (ret)
 765                                 goto err;
 766
 767                         if (ctxt->stats)
 768                                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 769                 } else {
 770                         struct btree *b;
 771
 772                         b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
 773                         ret = PTR_ERR_OR_ZERO(b);
 774                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 775                                 continue;
 776                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 777                                 continue;
 778                         if (ret)
 779                                 goto err;
 780                         if (!b)
 781                                 goto next;
 782
 783                         unsigned sectors = btree_ptr_sectors_written(bkey_i_to_s_c(&b->key));
 784
 785                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 786                         bch2_trans_iter_exit(trans, &iter);
 787
 788                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 789                                 continue;
 790                         if (ret)
 791                                 goto err;
 792
 793                         if (ctxt->rate)
 794                                 bch2_ratelimit_increment(ctxt->rate, sectors);
 795                         if (ctxt->stats) {
 796                                 atomic64_add(sectors, &ctxt->stats->sectors_seen);
 797                                 atomic64_add(sectors, &ctxt->stats->sectors_moved);
 798                         }
 799                 }
 800 next:
 801                 bp_pos = bpos_nosnap_successor(bp_pos);
 802         }
 803
 804         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 805 err:
 806         bch2_dev_put(ca);
 807         bch2_bkey_buf_exit(&sk, c);
 808         return ret;
 809 }
 810
 811 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 812                                 struct btree *, struct bch_io_opts *,
 813                                 struct data_update_opts *);
 814
 815 static int bch2_move_btree(struct bch_fs *c,
 816                            struct bbpos start,
 817                            struct bbpos end,
 818                            move_btree_pred pred, void *arg,
 819                            struct bch_move_stats *stats)
 820 {
 821         bool kthread = (current->flags & PF_KTHREAD) != 0;
 822         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 823         struct moving_context ctxt;
 824         struct btree_trans *trans;
 825         struct btree_iter iter;
 826         struct btree *b;
 827         enum btree_id btree;
 828         struct data_update_opts data_opts;
 829         int ret = 0;
 830
 831         bch2_moving_ctxt_init(&ctxt, c, NULL, stats,
 832                               writepoint_ptr(&c->btree_write_point),
 833                               true);
 834         trans = ctxt.trans;
 835
 836         stats->data_type = BCH_DATA_btree;
 837
 838         for (btree = start.btree;
 839              btree <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
 840              btree ++) {
 841                 stats->pos = BBPOS(btree, POS_MIN);
 842
 843                 if (!bch2_btree_id_root(c, btree)->b)
 844                         continue;
 845
 846                 bch2_trans_node_iter_init(trans, &iter, btree, POS_MIN, 0, 0,
 847                                           BTREE_ITER_prefetch);
 848 retry:
 849                 ret = 0;
 850                 while (bch2_trans_begin(trans),
 851                        (b = bch2_btree_iter_peek_node(&iter)) &&
 852                        !(ret = PTR_ERR_OR_ZERO(b))) {
 853                         if (kthread && kthread_should_stop())
 854                                 break;
 855
 856                         if ((cmp_int(btree, end.btree) ?:
 857                              bpos_cmp(b->key.k.p, end.pos)) > 0)
 858                                 break;
 859
 860                         stats->pos = BBPOS(iter.btree_id, iter.pos);
 861
 862                         if (!pred(c, arg, b, &io_opts, &data_opts))
 863                                 goto next;
 864
 865                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0) ?: ret;
 866                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 867                                 continue;
 868                         if (ret)
 869                                 break;
 870 next:
 871                         bch2_btree_iter_next_node(&iter);
 872                 }
 873                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 874                         goto retry;
 875
 876                 bch2_trans_iter_exit(trans, &iter);
 877
 878                 if (kthread && kthread_should_stop())
 879                         break;
 880         }
 881
 882         bch_err_fn(c, ret);
 883         bch2_moving_ctxt_exit(&ctxt);
 884         bch2_btree_interior_updates_flush(c);
 885
 886         return ret;
 887 }
 888
 889 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 890                              struct bkey_s_c k,
 891                              struct bch_io_opts *io_opts,
 892                              struct data_update_opts *data_opts)
 893 {
 894         unsigned nr_good = bch2_bkey_durability(c, k);
 895         unsigned replicas = bkey_is_btree_ptr(k.k)
 896                 ? c->opts.metadata_replicas
 897                 : io_opts->data_replicas;
 898
 899         rcu_read_lock();
 900         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 901         unsigned i = 0;
 902         bkey_for_each_ptr(ptrs, ptr) {
 903                 struct bch_dev *ca = bch2_dev_rcu(c, ptr->dev);
 904                 if (!ptr->cached &&
 905                     (!ca || !ca->mi.durability))
 906                         data_opts->kill_ptrs |= BIT(i);
 907                 i++;
 908         }
 909         rcu_read_unlock();
 910
 911         if (!data_opts->kill_ptrs &&
 912             (!nr_good || nr_good >= replicas))
 913                 return false;
 914
 915         data_opts->target               = 0;
 916         data_opts->extra_replicas       = replicas - nr_good;
 917         data_opts->btree_insert_flags   = 0;
 918         return true;
 919 }
 920
 921 static bool migrate_pred(struct bch_fs *c, void *arg,
 922                          struct bkey_s_c k,
 923                          struct bch_io_opts *io_opts,
 924                          struct data_update_opts *data_opts)
 925 {
 926         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 927         struct bch_ioctl_data *op = arg;
 928         unsigned i = 0;
 929
 930         data_opts->rewrite_ptrs         = 0;
 931         data_opts->target               = 0;
 932         data_opts->extra_replicas       = 0;
 933         data_opts->btree_insert_flags   = 0;
 934
 935         bkey_for_each_ptr(ptrs, ptr) {
 936                 if (ptr->dev == op->migrate.dev)
 937                         data_opts->rewrite_ptrs |= 1U << i;
 938                 i++;
 939         }
 940
 941         return data_opts->rewrite_ptrs != 0;
 942 }
 943
 944 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 945                                    struct btree *b,
 946                                    struct bch_io_opts *io_opts,
 947                                    struct data_update_opts *data_opts)
 948 {
 949         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 950 }
 951
 952 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 953                                struct btree *b,
 954                                struct bch_io_opts *io_opts,
 955                                struct data_update_opts *data_opts)
 956 {
 957         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 958 }
 959
 960 /*
 961  * Ancient versions of bcachefs produced packed formats which could represent
 962  * keys that the in memory format cannot represent; this checks for those
 963  * formats so we can get rid of them.
 964  */
 965 static bool bformat_needs_redo(struct bkey_format *f)
 966 {
 967         for (unsigned i = 0; i < f->nr_fields; i++)
 968                 if (bch2_bkey_format_field_overflows(f, i))
 969                         return true;
 970
 971         return false;
 972 }
 973
 974 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 975                                    struct btree *b,
 976                                    struct bch_io_opts *io_opts,
 977                                    struct data_update_opts *data_opts)
 978 {
 979         if (b->version_ondisk != c->sb.version ||
 980             btree_node_need_rewrite(b) ||
 981             bformat_needs_redo(&b->format)) {
 982                 data_opts->target               = 0;
 983                 data_opts->extra_replicas       = 0;
 984                 data_opts->btree_insert_flags   = 0;
 985                 return true;
 986         }
 987
 988         return false;
 989 }
 990
 991 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
 992 {
 993         int ret;
 994
 995         ret = bch2_move_btree(c,
 996                               BBPOS_MIN,
 997                               BBPOS_MAX,
 998                               rewrite_old_nodes_pred, c, stats);
 999         if (!ret) {
1000                 mutex_lock(&c->sb_lock);
1001                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1002                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1003                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1004                 bch2_write_super(c);
1005                 mutex_unlock(&c->sb_lock);
1006         }
1007
1008         bch_err_fn(c, ret);
1009         return ret;
1010 }
1011
1012 static bool drop_extra_replicas_pred(struct bch_fs *c, void *arg,
1013                              struct bkey_s_c k,
1014                              struct bch_io_opts *io_opts,
1015                              struct data_update_opts *data_opts)
1016 {
1017         unsigned durability = bch2_bkey_durability(c, k);
1018         unsigned replicas = bkey_is_btree_ptr(k.k)
1019                 ? c->opts.metadata_replicas
1020                 : io_opts->data_replicas;
1021         const union bch_extent_entry *entry;
1022         struct extent_ptr_decoded p;
1023         unsigned i = 0;
1024
1025         rcu_read_lock();
1026         bkey_for_each_ptr_decode(k.k, bch2_bkey_ptrs_c(k), p, entry) {
1027                 unsigned d = bch2_extent_ptr_durability(c, &p);
1028
1029                 if (d && durability - d >= replicas) {
1030                         data_opts->kill_ptrs |= BIT(i);
1031                         durability -= d;
1032                 }
1033
1034                 i++;
1035         }
1036         rcu_read_unlock();
1037
1038         return data_opts->kill_ptrs != 0;
1039 }
1040
1041 static bool drop_extra_replicas_btree_pred(struct bch_fs *c, void *arg,
1042                                    struct btree *b,
1043                                    struct bch_io_opts *io_opts,
1044                                    struct data_update_opts *data_opts)
1045 {
1046         return drop_extra_replicas_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
1047 }
1048
1049 int bch2_data_job(struct bch_fs *c,
1050                   struct bch_move_stats *stats,
1051                   struct bch_ioctl_data op)
1052 {
1053         struct bbpos start      = BBPOS(op.start_btree, op.start_pos);
1054         struct bbpos end        = BBPOS(op.end_btree, op.end_pos);
1055         int ret = 0;
1056
1057         if (op.op >= BCH_DATA_OP_NR)
1058                 return -EINVAL;
1059
1060         bch2_move_stats_init(stats, bch2_data_ops_strs[op.op]);
1061
1062         switch (op.op) {
1063         case BCH_DATA_OP_rereplicate:
1064                 stats->data_type = BCH_DATA_journal;
1065                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1066                 ret = bch2_move_btree(c, start, end,
1067                                       rereplicate_btree_pred, c, stats) ?: ret;
1068                 ret = bch2_move_data(c, start, end,
1069                                      NULL,
1070                                      stats,
1071                                      writepoint_hashed((unsigned long) current),
1072                                      true,
1073                                      rereplicate_pred, c) ?: ret;
1074                 ret = bch2_replicas_gc2(c) ?: ret;
1075                 break;
1076         case BCH_DATA_OP_migrate:
1077                 if (op.migrate.dev >= c->sb.nr_devices)
1078                         return -EINVAL;
1079
1080                 stats->data_type = BCH_DATA_journal;
1081                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1082                 ret = bch2_move_btree(c, start, end,
1083                                       migrate_btree_pred, &op, stats) ?: ret;
1084                 ret = bch2_move_data(c, start, end,
1085                                      NULL,
1086                                      stats,
1087                                      writepoint_hashed((unsigned long) current),
1088                                      true,
1089                                      migrate_pred, &op) ?: ret;
1090                 ret = bch2_replicas_gc2(c) ?: ret;
1091                 break;
1092         case BCH_DATA_OP_rewrite_old_nodes:
1093                 ret = bch2_scan_old_btree_nodes(c, stats);
1094                 break;
1095         case BCH_DATA_OP_drop_extra_replicas:
1096                 ret = bch2_move_btree(c, start, end,
1097                                 drop_extra_replicas_btree_pred, c, stats) ?: ret;
1098                 ret = bch2_move_data(c, start, end, NULL, stats,
1099                                 writepoint_hashed((unsigned long) current),
1100                                 true,
1101                                 drop_extra_replicas_pred, c) ?: ret;
1102                 ret = bch2_replicas_gc2(c) ?: ret;
1103                 break;
1104         default:
1105                 ret = -EINVAL;
1106         }
1107
1108         bch2_move_stats_exit(stats, c);
1109         return ret;
1110 }
1111
1112 void bch2_move_stats_to_text(struct printbuf *out, struct bch_move_stats *stats)
1113 {
1114         prt_printf(out, "%s: data type==", stats->name);
1115         bch2_prt_data_type(out, stats->data_type);
1116         prt_str(out, " pos=");
1117         bch2_bbpos_to_text(out, stats->pos);
1118         prt_newline(out);
1119         printbuf_indent_add(out, 2);
1120
1121         prt_printf(out, "keys moved:  %llu\n",  atomic64_read(&stats->keys_moved));
1122         prt_printf(out, "keys raced:  %llu\n",  atomic64_read(&stats->keys_raced));
1123         prt_printf(out, "bytes seen:  ");
1124         prt_human_readable_u64(out, atomic64_read(&stats->sectors_seen) << 9);
1125         prt_newline(out);
1126
1127         prt_printf(out, "bytes moved: ");
1128         prt_human_readable_u64(out, atomic64_read(&stats->sectors_moved) << 9);
1129         prt_newline(out);
1130
1131         prt_printf(out, "bytes raced: ");
1132         prt_human_readable_u64(out, atomic64_read(&stats->sectors_raced) << 9);
1133         prt_newline(out);
1134
1135         printbuf_indent_sub(out, 2);
1136 }
1137
1138 static void bch2_moving_ctxt_to_text(struct printbuf *out, struct bch_fs *c, struct moving_context *ctxt)
1139 {
1140         struct moving_io *io;
1141
1142         bch2_move_stats_to_text(out, ctxt->stats);
1143         printbuf_indent_add(out, 2);
1144
1145         prt_printf(out, "reads: ios %u/%u sectors %u/%u\n",
1146                    atomic_read(&ctxt->read_ios),
1147                    c->opts.move_ios_in_flight,
1148                    atomic_read(&ctxt->read_sectors),
1149                    c->opts.move_bytes_in_flight >> 9);
1150
1151         prt_printf(out, "writes: ios %u/%u sectors %u/%u\n",
1152                    atomic_read(&ctxt->write_ios),
1153                    c->opts.move_ios_in_flight,
1154                    atomic_read(&ctxt->write_sectors),
1155                    c->opts.move_bytes_in_flight >> 9);
1156
1157         printbuf_indent_add(out, 2);
1158
1159         mutex_lock(&ctxt->lock);
1160         list_for_each_entry(io, &ctxt->ios, io_list)
1161                 bch2_write_op_to_text(out, &io->write.op);
1162         mutex_unlock(&ctxt->lock);
1163
1164         printbuf_indent_sub(out, 4);
1165 }
1166
1167 void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
1168 {
1169         struct moving_context *ctxt;
1170
1171         mutex_lock(&c->moving_context_lock);
1172         list_for_each_entry(ctxt, &c->moving_context_list, list)
1173                 bch2_moving_ctxt_to_text(out, c, ctxt);
1174         mutex_unlock(&c->moving_context_lock);
1175 }
1176
1177 void bch2_fs_move_init(struct bch_fs *c)
1178 {
1179         INIT_LIST_HEAD(&c->moving_context_list);
1180         mutex_init(&c->moving_context_lock);
1181 }