fs/bcachefs/data_update.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_buf.h"
   6 #include "btree_update.h"
   7 #include "buckets.h"
   8 #include "compress.h"
   9 #include "data_update.h"
  10 #include "disk_groups.h"
  11 #include "ec.h"
  12 #include "error.h"
  13 #include "extents.h"
  14 #include "io_write.h"
  15 #include "keylist.h"
  16 #include "move.h"
  17 #include "nocow_locking.h"
  18 #include "rebalance.h"
  19 #include "snapshot.h"
  20 #include "subvolume.h"
  21 #include "trace.h"
  22
  23 static void bkey_put_dev_refs(struct bch_fs *c, struct bkey_s_c k)
  24 {
  25         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  26
  27         bkey_for_each_ptr(ptrs, ptr)
  28                 bch2_dev_put(bch2_dev_have_ref(c, ptr->dev));
  29 }
  30
  31 static bool bkey_get_dev_refs(struct bch_fs *c, struct bkey_s_c k)
  32 {
  33         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  34
  35         bkey_for_each_ptr(ptrs, ptr) {
  36                 if (!bch2_dev_tryget(c, ptr->dev)) {
  37                         bkey_for_each_ptr(ptrs, ptr2) {
  38                                 if (ptr2 == ptr)
  39                                         break;
  40                                 bch2_dev_put(bch2_dev_have_ref(c, ptr2->dev));
  41                         }
  42                         return false;
  43                 }
  44         }
  45         return true;
  46 }
  47
  48 static void bkey_nocow_unlock(struct bch_fs *c, struct bkey_s_c k)
  49 {
  50         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  51
  52         bkey_for_each_ptr(ptrs, ptr) {
  53                 struct bch_dev *ca = bch2_dev_have_ref(c, ptr->dev);
  54                 struct bpos bucket = PTR_BUCKET_POS(ca, ptr);
  55
  56                 bch2_bucket_nocow_unlock(&c->nocow_locks, bucket, 0);
  57         }
  58 }
  59
  60 static bool bkey_nocow_lock(struct bch_fs *c, struct moving_context *ctxt, struct bkey_s_c k)
  61 {
  62         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  63
  64         bkey_for_each_ptr(ptrs, ptr) {
  65                 struct bch_dev *ca = bch2_dev_have_ref(c, ptr->dev);
  66                 struct bpos bucket = PTR_BUCKET_POS(ca, ptr);
  67
  68                 if (ctxt) {
  69                         bool locked;
  70
  71                         move_ctxt_wait_event(ctxt,
  72                                 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks, bucket, 0)) ||
  73                                 list_empty(&ctxt->ios));
  74
  75                         if (!locked)
  76                                 bch2_bucket_nocow_lock(&c->nocow_locks, bucket, 0);
  77                 } else {
  78                         if (!bch2_bucket_nocow_trylock(&c->nocow_locks, bucket, 0)) {
  79                                 bkey_for_each_ptr(ptrs, ptr2) {
  80                                         if (ptr2 == ptr)
  81                                                 break;
  82
  83                                         ca = bch2_dev_have_ref(c, ptr2->dev);
  84                                         bucket = PTR_BUCKET_POS(ca, ptr2);
  85                                         bch2_bucket_nocow_unlock(&c->nocow_locks, bucket, 0);
  86                                 }
  87                                 return false;
  88                         }
  89                 }
  90         }
  91         return true;
  92 }
  93
  94 static noinline void trace_move_extent_finish2(struct data_update *u,
  95                                                struct bkey_i *new,
  96                                                struct bkey_i *insert)
  97 {
  98         struct bch_fs *c = u->op.c;
  99         struct printbuf buf = PRINTBUF;
 100
 101         prt_newline(&buf);
 102
 103         bch2_data_update_to_text(&buf, u);
 104         prt_newline(&buf);
 105
 106         prt_str_indented(&buf, "new replicas:\t");
 107         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 108         prt_newline(&buf);
 109
 110         prt_str_indented(&buf, "insert:\t");
 111         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 112         prt_newline(&buf);
 113
 114         trace_move_extent_finish(c, buf.buf);
 115         printbuf_exit(&buf);
 116 }
 117
 118 static void trace_move_extent_fail2(struct data_update *m,
 119                          struct bkey_s_c new,
 120                          struct bkey_s_c wrote,
 121                          struct bkey_i *insert,
 122                          const char *msg)
 123 {
 124         struct bch_fs *c = m->op.c;
 125         struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 126         struct printbuf buf = PRINTBUF;
 127         unsigned rewrites_found = 0;
 128
 129         if (!trace_move_extent_fail_enabled())
 130                 return;
 131
 132         prt_str(&buf, msg);
 133
 134         if (insert) {
 135                 const union bch_extent_entry *entry;
 136                 struct bch_extent_ptr *ptr;
 137                 struct extent_ptr_decoded p;
 138
 139                 unsigned ptr_bit = 1;
 140                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
 141                         if ((ptr_bit & m->data_opts.rewrite_ptrs) &&
 142                             (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
 143                             !ptr->cached)
 144                                 rewrites_found |= ptr_bit;
 145                         ptr_bit <<= 1;
 146                 }
 147         }
 148
 149         prt_str(&buf, "rewrites found:\t");
 150         bch2_prt_u64_base2(&buf, rewrites_found);
 151         prt_newline(&buf);
 152
 153         bch2_data_update_opts_to_text(&buf, c, &m->op.opts, &m->data_opts);
 154
 155         prt_str(&buf, "\nold:    ");
 156         bch2_bkey_val_to_text(&buf, c, old);
 157
 158         prt_str(&buf, "\nnew:    ");
 159         bch2_bkey_val_to_text(&buf, c, new);
 160
 161         prt_str(&buf, "\nwrote:  ");
 162         bch2_bkey_val_to_text(&buf, c, wrote);
 163
 164         if (insert) {
 165                 prt_str(&buf, "\ninsert: ");
 166                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 167         }
 168
 169         trace_move_extent_fail(c, buf.buf);
 170         printbuf_exit(&buf);
 171 }
 172
 173 static int __bch2_data_update_index_update(struct btree_trans *trans,
 174                                            struct bch_write_op *op)
 175 {
 176         struct bch_fs *c = op->c;
 177         struct btree_iter iter;
 178         struct data_update *m =
 179                 container_of(op, struct data_update, op);
 180         struct keylist *keys = &op->insert_keys;
 181         struct bkey_buf _new, _insert;
 182         int ret = 0;
 183
 184         bch2_bkey_buf_init(&_new);
 185         bch2_bkey_buf_init(&_insert);
 186         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 187
 188         bch2_trans_iter_init(trans, &iter, m->btree_id,
 189                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 190                              BTREE_ITER_slots|BTREE_ITER_intent);
 191
 192         while (1) {
 193                 struct bkey_s_c k;
 194                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 195                 struct bkey_i *insert = NULL;
 196                 struct bkey_i_extent *new;
 197                 const union bch_extent_entry *entry_c;
 198                 union bch_extent_entry *entry;
 199                 struct extent_ptr_decoded p;
 200                 struct bch_extent_ptr *ptr;
 201                 const struct bch_extent_ptr *ptr_c;
 202                 struct bpos next_pos;
 203                 bool should_check_enospc;
 204                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 205                 unsigned rewrites_found = 0, durability, ptr_bit;
 206
 207                 bch2_trans_begin(trans);
 208
 209                 k = bch2_btree_iter_peek_slot(&iter);
 210                 ret = bkey_err(k);
 211                 if (ret)
 212                         goto err;
 213
 214                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 215
 216                 if (!bch2_extents_match(k, old)) {
 217                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i),
 218                                                 NULL, "no match:");
 219                         goto nowork;
 220                 }
 221
 222                 bkey_reassemble(_insert.k, k);
 223                 insert = _insert.k;
 224
 225                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 226                 new = bkey_i_to_extent(_new.k);
 227                 bch2_cut_front(iter.pos, &new->k_i);
 228
 229                 bch2_cut_front(iter.pos,        insert);
 230                 bch2_cut_back(new->k.p,         insert);
 231                 bch2_cut_back(insert->k.p,      &new->k_i);
 232
 233                 /*
 234                  * @old: extent that we read from
 235                  * @insert: key that we're going to update, initialized from
 236                  * extent currently in btree - same as @old unless we raced with
 237                  * other updates
 238                  * @new: extent with new pointers that we'll be adding to @insert
 239                  *
 240                  * Fist, drop rewrite_ptrs from @new:
 241                  */
 242                 ptr_bit = 1;
 243                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
 244                         if ((ptr_bit & m->data_opts.rewrite_ptrs) &&
 245                             (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
 246                             !ptr->cached) {
 247                                 bch2_extent_ptr_set_cached(c, &m->op.opts,
 248                                                            bkey_i_to_s(insert), ptr);
 249                                 rewrites_found |= ptr_bit;
 250                         }
 251                         ptr_bit <<= 1;
 252                 }
 253
 254                 if (m->data_opts.rewrite_ptrs &&
 255                     !rewrites_found &&
 256                     bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) {
 257                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:");
 258                         goto nowork;
 259                 }
 260
 261                 /*
 262                  * A replica that we just wrote might conflict with a replica
 263                  * that we want to keep, due to racing with another move:
 264                  */
 265 restart_drop_conflicting_replicas:
 266                 extent_for_each_ptr(extent_i_to_s(new), ptr)
 267                         if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
 268                             !ptr_c->cached) {
 269                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
 270                                 goto restart_drop_conflicting_replicas;
 271                         }
 272
 273                 if (!bkey_val_u64s(&new->k)) {
 274                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:");
 275                         goto nowork;
 276                 }
 277
 278                 /* Now, drop pointers that conflict with what we just wrote: */
 279                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
 280                         if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev)))
 281                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
 282
 283                 durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) +
 284                         bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i));
 285
 286                 /* Now, drop excess replicas: */
 287                 rcu_read_lock();
 288 restart_drop_extra_replicas:
 289                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
 290                         unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
 291
 292                         if (!p.ptr.cached &&
 293                             durability - ptr_durability >= m->op.opts.data_replicas) {
 294                                 durability -= ptr_durability;
 295
 296                                 bch2_extent_ptr_set_cached(c, &m->op.opts,
 297                                                            bkey_i_to_s(insert), &entry->ptr);
 298                                 goto restart_drop_extra_replicas;
 299                         }
 300                 }
 301                 rcu_read_unlock();
 302
 303                 /* Finally, add the pointers we just wrote: */
 304                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
 305                         bch2_extent_ptr_decoded_append(insert, &p);
 306
 307                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 308                 bch2_extent_normalize_by_opts(c, &m->op.opts, bkey_i_to_s(insert));
 309
 310                 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
 311                                                  &should_check_enospc,
 312                                                  &i_sectors_delta,
 313                                                  &disk_sectors_delta);
 314                 if (ret)
 315                         goto err;
 316
 317                 if (disk_sectors_delta > (s64) op->res.sectors) {
 318                         ret = bch2_disk_reservation_add(c, &op->res,
 319                                                 disk_sectors_delta - op->res.sectors,
 320                                                 !should_check_enospc
 321                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 322                         if (ret)
 323                                 goto out;
 324                 }
 325
 326                 next_pos = insert->k.p;
 327
 328                 /*
 329                  * Check for nonce offset inconsistency:
 330                  * This is debug code - we've been seeing this bug rarely, and
 331                  * it's been hard to reproduce, so this should give us some more
 332                  * information when it does occur:
 333                  */
 334                 int invalid = bch2_bkey_validate(c, bkey_i_to_s_c(insert),
 335                                                  (struct bkey_validate_context) {
 336                                                         .btree  = m->btree_id,
 337                                                         .flags  = BCH_VALIDATE_commit,
 338                                                  });
 339                 if (invalid) {
 340                         struct printbuf buf = PRINTBUF;
 341
 342                         prt_str(&buf, "about to insert invalid key in data update path");
 343                         prt_str(&buf, "\nold: ");
 344                         bch2_bkey_val_to_text(&buf, c, old);
 345                         prt_str(&buf, "\nk:   ");
 346                         bch2_bkey_val_to_text(&buf, c, k);
 347                         prt_str(&buf, "\nnew: ");
 348                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 349
 350                         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 351                         printbuf_exit(&buf);
 352
 353                         bch2_fatal_error(c);
 354                         ret = -EIO;
 355                         goto out;
 356                 }
 357
 358                 if (trace_data_update_enabled()) {
 359                         struct printbuf buf = PRINTBUF;
 360
 361                         prt_str(&buf, "\nold: ");
 362                         bch2_bkey_val_to_text(&buf, c, old);
 363                         prt_str(&buf, "\nk:   ");
 364                         bch2_bkey_val_to_text(&buf, c, k);
 365                         prt_str(&buf, "\nnew: ");
 366                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 367
 368                         trace_data_update(c, buf.buf);
 369                         printbuf_exit(&buf);
 370                 }
 371
 372                 ret =   bch2_insert_snapshot_whiteouts(trans, m->btree_id,
 373                                                 k.k->p, bkey_start_pos(&insert->k)) ?:
 374                         bch2_insert_snapshot_whiteouts(trans, m->btree_id,
 375                                                 k.k->p, insert->k.p) ?:
 376                         bch2_bkey_set_needs_rebalance(c, &op->opts, insert) ?:
 377                         bch2_trans_update(trans, &iter, insert,
 378                                 BTREE_UPDATE_internal_snapshot_node) ?:
 379                         bch2_trans_commit(trans, &op->res,
 380                                 NULL,
 381                                 BCH_TRANS_COMMIT_no_check_rw|
 382                                 BCH_TRANS_COMMIT_no_enospc|
 383                                 m->data_opts.btree_insert_flags);
 384                 if (!ret) {
 385                         bch2_btree_iter_set_pos(&iter, next_pos);
 386
 387                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 388                         if (trace_move_extent_finish_enabled())
 389                                 trace_move_extent_finish2(m, &new->k_i, insert);
 390                 }
 391 err:
 392                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 393                         ret = 0;
 394                 if (ret)
 395                         break;
 396 next:
 397                 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
 398                         bch2_keylist_pop_front(keys);
 399                         if (bch2_keylist_empty(keys))
 400                                 goto out;
 401                 }
 402                 continue;
 403 nowork:
 404                 if (m->stats) {
 405                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 406                         atomic64_inc(&m->stats->keys_raced);
 407                         atomic64_add(k.k->p.offset - iter.pos.offset,
 408                                      &m->stats->sectors_raced);
 409                 }
 410
 411                 count_event(c, move_extent_fail);
 412
 413                 bch2_btree_iter_advance(&iter);
 414                 goto next;
 415         }
 416 out:
 417         bch2_trans_iter_exit(trans, &iter);
 418         bch2_bkey_buf_exit(&_insert, c);
 419         bch2_bkey_buf_exit(&_new, c);
 420         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 421         return ret;
 422 }
 423
 424 int bch2_data_update_index_update(struct bch_write_op *op)
 425 {
 426         return bch2_trans_run(op->c, __bch2_data_update_index_update(trans, op));
 427 }
 428
 429 void bch2_data_update_read_done(struct data_update *m,
 430                                 struct bch_extent_crc_unpacked crc)
 431 {
 432         /* write bio must own pages: */
 433         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 434
 435         m->op.crc = crc;
 436         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 437
 438         closure_call(&m->op.cl, bch2_write, NULL, NULL);
 439 }
 440
 441 void bch2_data_update_exit(struct data_update *update)
 442 {
 443         struct bch_fs *c = update->op.c;
 444         struct bkey_s_c k = bkey_i_to_s_c(update->k.k);
 445
 446         if (c->opts.nocow_enabled)
 447                 bkey_nocow_unlock(c, k);
 448         bkey_put_dev_refs(c, k);
 449         bch2_bkey_buf_exit(&update->k, c);
 450         bch2_disk_reservation_put(c, &update->op.res);
 451         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 452 }
 453
 454 static void bch2_update_unwritten_extent(struct btree_trans *trans,
 455                                   struct data_update *update)
 456 {
 457         struct bch_fs *c = update->op.c;
 458         struct bio *bio = &update->op.wbio.bio;
 459         struct bkey_i_extent *e;
 460         struct write_point *wp;
 461         struct closure cl;
 462         struct btree_iter iter;
 463         struct bkey_s_c k;
 464         int ret;
 465
 466         closure_init_stack(&cl);
 467         bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
 468
 469         while (bio_sectors(bio)) {
 470                 unsigned sectors = bio_sectors(bio);
 471
 472                 bch2_trans_begin(trans);
 473
 474                 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
 475                                      BTREE_ITER_slots);
 476                 ret = lockrestart_do(trans, ({
 477                         k = bch2_btree_iter_peek_slot(&iter);
 478                         bkey_err(k);
 479                 }));
 480                 bch2_trans_iter_exit(trans, &iter);
 481
 482                 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
 483                         break;
 484
 485                 e = bkey_extent_init(update->op.insert_keys.top);
 486                 e->k.p = update->op.pos;
 487
 488                 ret = bch2_alloc_sectors_start_trans(trans,
 489                                 update->op.target,
 490                                 false,
 491                                 update->op.write_point,
 492                                 &update->op.devs_have,
 493                                 update->op.nr_replicas,
 494                                 update->op.nr_replicas,
 495                                 update->op.watermark,
 496                                 0, &cl, &wp);
 497                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
 498                         bch2_trans_unlock(trans);
 499                         closure_sync(&cl);
 500                         continue;
 501                 }
 502
 503                 bch_err_fn_ratelimited(c, ret);
 504
 505                 if (ret)
 506                         return;
 507
 508                 sectors = min(sectors, wp->sectors_free);
 509
 510                 bch2_key_resize(&e->k, sectors);
 511
 512                 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
 513                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 514                 bch2_alloc_sectors_done(c, wp);
 515
 516                 bio_advance(bio, sectors << 9);
 517                 update->op.pos.offset += sectors;
 518
 519                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 520                         ptr->unwritten = true;
 521                 bch2_keylist_push(&update->op.insert_keys);
 522
 523                 ret = __bch2_data_update_index_update(trans, &update->op);
 524
 525                 bch2_open_buckets_put(c, &update->op.open_buckets);
 526
 527                 if (ret)
 528                         break;
 529         }
 530
 531         if (closure_nr_remaining(&cl) != 1) {
 532                 bch2_trans_unlock(trans);
 533                 closure_sync(&cl);
 534         }
 535 }
 536
 537 void bch2_data_update_opts_to_text(struct printbuf *out, struct bch_fs *c,
 538                                    struct bch_io_opts *io_opts,
 539                                    struct data_update_opts *data_opts)
 540 {
 541         printbuf_tabstop_push(out, 20);
 542
 543         prt_str_indented(out, "rewrite ptrs:\t");
 544         bch2_prt_u64_base2(out, data_opts->rewrite_ptrs);
 545         prt_newline(out);
 546
 547         prt_str_indented(out, "kill ptrs:\t");
 548         bch2_prt_u64_base2(out, data_opts->kill_ptrs);
 549         prt_newline(out);
 550
 551         prt_str_indented(out, "target:\t");
 552         bch2_target_to_text(out, c, data_opts->target);
 553         prt_newline(out);
 554
 555         prt_str_indented(out, "compression:\t");
 556         bch2_compression_opt_to_text(out, io_opts->background_compression);
 557         prt_newline(out);
 558
 559         prt_str_indented(out, "opts.replicas:\t");
 560         prt_u64(out, io_opts->data_replicas);
 561         prt_newline(out);
 562
 563         prt_str_indented(out, "extra replicas:\t");
 564         prt_u64(out, data_opts->extra_replicas);
 565 }
 566
 567 void bch2_data_update_to_text(struct printbuf *out, struct data_update *m)
 568 {
 569         bch2_data_update_opts_to_text(out, m->op.c, &m->op.opts, &m->data_opts);
 570         prt_newline(out);
 571
 572         prt_str_indented(out, "old key:\t");
 573         bch2_bkey_val_to_text(out, m->op.c, bkey_i_to_s_c(m->k.k));
 574 }
 575
 576 int bch2_extent_drop_ptrs(struct btree_trans *trans,
 577                           struct btree_iter *iter,
 578                           struct bkey_s_c k,
 579                           struct bch_io_opts *io_opts,
 580                           struct data_update_opts *data_opts)
 581 {
 582         struct bch_fs *c = trans->c;
 583         struct bkey_i *n;
 584         int ret;
 585
 586         n = bch2_bkey_make_mut_noupdate(trans, k);
 587         ret = PTR_ERR_OR_ZERO(n);
 588         if (ret)
 589                 return ret;
 590
 591         while (data_opts->kill_ptrs) {
 592                 unsigned i = 0, drop = __fls(data_opts->kill_ptrs);
 593
 594                 bch2_bkey_drop_ptrs_noerror(bkey_i_to_s(n), ptr, i++ == drop);
 595                 data_opts->kill_ptrs ^= 1U << drop;
 596         }
 597
 598         /*
 599          * If the new extent no longer has any pointers, bch2_extent_normalize()
 600          * will do the appropriate thing with it (turning it into a
 601          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 602          */
 603         bch2_extent_normalize_by_opts(c, io_opts, bkey_i_to_s(n));
 604
 605         /*
 606          * Since we're not inserting through an extent iterator
 607          * (BTREE_ITER_all_snapshots iterators aren't extent iterators),
 608          * we aren't using the extent overwrite path to delete, we're
 609          * just using the normal key deletion path:
 610          */
 611         if (bkey_deleted(&n->k) && !(iter->flags & BTREE_ITER_is_extents))
 612                 n->k.size = 0;
 613
 614         return bch2_trans_relock(trans) ?:
 615                 bch2_trans_update(trans, iter, n, BTREE_UPDATE_internal_snapshot_node) ?:
 616                 bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
 617 }
 618
 619 int bch2_data_update_init(struct btree_trans *trans,
 620                           struct btree_iter *iter,
 621                           struct moving_context *ctxt,
 622                           struct data_update *m,
 623                           struct write_point_specifier wp,
 624                           struct bch_io_opts io_opts,
 625                           struct data_update_opts data_opts,
 626                           enum btree_id btree_id,
 627                           struct bkey_s_c k)
 628 {
 629         struct bch_fs *c = trans->c;
 630         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 631         const union bch_extent_entry *entry;
 632         struct extent_ptr_decoded p;
 633         unsigned reserve_sectors = k.k->size * data_opts.extra_replicas;
 634         int ret = 0;
 635
 636         /*
 637          * fs is corrupt  we have a key for a snapshot node that doesn't exist,
 638          * and we have to check for this because we go rw before repairing the
 639          * snapshots table - just skip it, we can move it later.
 640          */
 641         if (unlikely(k.k->p.snapshot && !bch2_snapshot_exists(c, k.k->p.snapshot)))
 642                 return -BCH_ERR_data_update_done;
 643
 644         if (!bkey_get_dev_refs(c, k))
 645                 return -BCH_ERR_data_update_done;
 646
 647         if (c->opts.nocow_enabled &&
 648             !bkey_nocow_lock(c, ctxt, k)) {
 649                 bkey_put_dev_refs(c, k);
 650                 return -BCH_ERR_nocow_lock_blocked;
 651         }
 652
 653         bch2_bkey_buf_init(&m->k);
 654         bch2_bkey_buf_reassemble(&m->k, c, k);
 655         m->btree_id     = btree_id;
 656         m->data_opts    = data_opts;
 657         m->ctxt         = ctxt;
 658         m->stats        = ctxt ? ctxt->stats : NULL;
 659
 660         bch2_write_op_init(&m->op, c, io_opts);
 661         m->op.pos       = bkey_start_pos(k.k);
 662         m->op.version   = k.k->bversion;
 663         m->op.target    = data_opts.target;
 664         m->op.write_point = wp;
 665         m->op.nr_replicas = 0;
 666         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 667                 BCH_WRITE_PAGES_OWNED|
 668                 BCH_WRITE_DATA_ENCODED|
 669                 BCH_WRITE_MOVE|
 670                 m->data_opts.write_flags;
 671         m->op.compression_opt   = io_opts.background_compression;
 672         m->op.watermark         = m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
 673
 674         unsigned durability_have = 0, durability_removing = 0;
 675
 676         unsigned ptr_bit = 1;
 677         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 678                 if (!p.ptr.cached) {
 679                         rcu_read_lock();
 680                         if (ptr_bit & m->data_opts.rewrite_ptrs) {
 681                                 if (crc_is_compressed(p.crc))
 682                                         reserve_sectors += k.k->size;
 683
 684                                 m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
 685                                 durability_removing += bch2_extent_ptr_desired_durability(c, &p);
 686                         } else if (!(ptr_bit & m->data_opts.kill_ptrs)) {
 687                                 bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 688                                 durability_have += bch2_extent_ptr_durability(c, &p);
 689                         }
 690                         rcu_read_unlock();
 691                 }
 692
 693                 /*
 694                  * op->csum_type is normally initialized from the fs/file's
 695                  * current options - but if an extent is encrypted, we require
 696                  * that it stays encrypted:
 697                  */
 698                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 699                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 700                         m->op.csum_type = p.crc.csum_type;
 701                 }
 702
 703                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 704                         m->op.incompressible = true;
 705
 706                 ptr_bit <<= 1;
 707         }
 708
 709         unsigned durability_required = max(0, (int) (io_opts.data_replicas - durability_have));
 710
 711         /*
 712          * If current extent durability is less than io_opts.data_replicas,
 713          * we're not trying to rereplicate the extent up to data_replicas here -
 714          * unless extra_replicas was specified
 715          *
 716          * Increasing replication is an explicit operation triggered by
 717          * rereplicate, currently, so that users don't get an unexpected -ENOSPC
 718          */
 719         m->op.nr_replicas = min(durability_removing, durability_required) +
 720                 m->data_opts.extra_replicas;
 721
 722         /*
 723          * If device(s) were set to durability=0 after data was written to them
 724          * we can end up with a duribilty=0 extent, and the normal algorithm
 725          * that tries not to increase durability doesn't work:
 726          */
 727         if (!(durability_have + durability_removing))
 728                 m->op.nr_replicas = max((unsigned) m->op.nr_replicas, 1);
 729
 730         m->op.nr_replicas_required = m->op.nr_replicas;
 731
 732         /*
 733          * It might turn out that we don't need any new replicas, if the
 734          * replicas or durability settings have been changed since the extent
 735          * was written:
 736          */
 737         if (!m->op.nr_replicas) {
 738                 m->data_opts.kill_ptrs |= m->data_opts.rewrite_ptrs;
 739                 m->data_opts.rewrite_ptrs = 0;
 740                 /* if iter == NULL, it's just a promote */
 741                 if (iter)
 742                         ret = bch2_extent_drop_ptrs(trans, iter, k, &io_opts, &m->data_opts);
 743                 goto out;
 744         }
 745
 746         if (reserve_sectors) {
 747                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 748                                 m->data_opts.extra_replicas
 749                                 ? 0
 750                                 : BCH_DISK_RESERVATION_NOFAIL);
 751                 if (ret)
 752                         goto out;
 753         }
 754
 755         if (bkey_extent_is_unwritten(k)) {
 756                 bch2_update_unwritten_extent(trans, m);
 757                 goto out;
 758         }
 759
 760         return 0;
 761 out:
 762         bch2_data_update_exit(m);
 763         return ret ?: -BCH_ERR_data_update_done;
 764 }
 765
 766 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 767 {
 768         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 769         unsigned ptr_bit = 1;
 770
 771         bkey_for_each_ptr(ptrs, ptr) {
 772                 if ((opts->rewrite_ptrs & ptr_bit) && ptr->cached) {
 773                         opts->kill_ptrs |= ptr_bit;
 774                         opts->rewrite_ptrs ^= ptr_bit;
 775                 }
 776
 777                 ptr_bit <<= 1;
 778         }
 779 }