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 "data_update.h"
   9 #include "ec.h"
  10 #include "error.h"
  11 #include "extents.h"
  12 #include "io_write.h"
  13 #include "keylist.h"
  14 #include "move.h"
  15 #include "nocow_locking.h"
  16 #include "rebalance.h"
  17 #include "snapshot.h"
  18 #include "subvolume.h"
  19 #include "trace.h"
  20
  21 static void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k)
  22 {
  23         if (trace_move_extent_finish_enabled()) {
  24                 struct printbuf buf = PRINTBUF;
  25
  26                 bch2_bkey_val_to_text(&buf, c, k);
  27                 trace_move_extent_finish(c, buf.buf);
  28                 printbuf_exit(&buf);
  29         }
  30 }
  31
  32 static void trace_move_extent_fail2(struct data_update *m,
  33                          struct bkey_s_c new,
  34                          struct bkey_s_c wrote,
  35                          struct bkey_i *insert,
  36                          const char *msg)
  37 {
  38         struct bch_fs *c = m->op.c;
  39         struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
  40         const union bch_extent_entry *entry;
  41         struct bch_extent_ptr *ptr;
  42         struct extent_ptr_decoded p;
  43         struct printbuf buf = PRINTBUF;
  44         unsigned i, rewrites_found = 0;
  45
  46         if (!trace_move_extent_fail_enabled())
  47                 return;
  48
  49         prt_str(&buf, msg);
  50
  51         if (insert) {
  52                 i = 0;
  53                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
  54                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
  55                             (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
  56                             !ptr->cached)
  57                                 rewrites_found |= 1U << i;
  58                         i++;
  59                 }
  60         }
  61
  62         prt_printf(&buf, "\nrewrite ptrs:   %u%u%u%u",
  63                    (m->data_opts.rewrite_ptrs & (1 << 0)) != 0,
  64                    (m->data_opts.rewrite_ptrs & (1 << 1)) != 0,
  65                    (m->data_opts.rewrite_ptrs & (1 << 2)) != 0,
  66                    (m->data_opts.rewrite_ptrs & (1 << 3)) != 0);
  67
  68         prt_printf(&buf, "\nrewrites found: %u%u%u%u",
  69                    (rewrites_found & (1 << 0)) != 0,
  70                    (rewrites_found & (1 << 1)) != 0,
  71                    (rewrites_found & (1 << 2)) != 0,
  72                    (rewrites_found & (1 << 3)) != 0);
  73
  74         prt_str(&buf, "\nold:    ");
  75         bch2_bkey_val_to_text(&buf, c, old);
  76
  77         prt_str(&buf, "\nnew:    ");
  78         bch2_bkey_val_to_text(&buf, c, new);
  79
  80         prt_str(&buf, "\nwrote:  ");
  81         bch2_bkey_val_to_text(&buf, c, wrote);
  82
  83         if (insert) {
  84                 prt_str(&buf, "\ninsert: ");
  85                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
  86         }
  87
  88         trace_move_extent_fail(c, buf.buf);
  89         printbuf_exit(&buf);
  90 }
  91
  92 static int __bch2_data_update_index_update(struct btree_trans *trans,
  93                                            struct bch_write_op *op)
  94 {
  95         struct bch_fs *c = op->c;
  96         struct btree_iter iter;
  97         struct data_update *m =
  98                 container_of(op, struct data_update, op);
  99         struct keylist *keys = &op->insert_keys;
 100         struct bkey_buf _new, _insert;
 101         int ret = 0;
 102
 103         bch2_bkey_buf_init(&_new);
 104         bch2_bkey_buf_init(&_insert);
 105         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 106
 107         bch2_trans_iter_init(trans, &iter, m->btree_id,
 108                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 109                              BTREE_ITER_slots|BTREE_ITER_intent);
 110
 111         while (1) {
 112                 struct bkey_s_c k;
 113                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 114                 struct bkey_i *insert = NULL;
 115                 struct bkey_i_extent *new;
 116                 const union bch_extent_entry *entry_c;
 117                 union bch_extent_entry *entry;
 118                 struct extent_ptr_decoded p;
 119                 struct bch_extent_ptr *ptr;
 120                 const struct bch_extent_ptr *ptr_c;
 121                 struct bpos next_pos;
 122                 bool should_check_enospc;
 123                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 124                 unsigned rewrites_found = 0, durability, i;
 125
 126                 bch2_trans_begin(trans);
 127
 128                 k = bch2_btree_iter_peek_slot(&iter);
 129                 ret = bkey_err(k);
 130                 if (ret)
 131                         goto err;
 132
 133                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 134
 135                 if (!bch2_extents_match(k, old)) {
 136                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i),
 137                                                 NULL, "no match:");
 138                         goto nowork;
 139                 }
 140
 141                 bkey_reassemble(_insert.k, k);
 142                 insert = _insert.k;
 143
 144                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 145                 new = bkey_i_to_extent(_new.k);
 146                 bch2_cut_front(iter.pos, &new->k_i);
 147
 148                 bch2_cut_front(iter.pos,        insert);
 149                 bch2_cut_back(new->k.p,         insert);
 150                 bch2_cut_back(insert->k.p,      &new->k_i);
 151
 152                 /*
 153                  * @old: extent that we read from
 154                  * @insert: key that we're going to update, initialized from
 155                  * extent currently in btree - same as @old unless we raced with
 156                  * other updates
 157                  * @new: extent with new pointers that we'll be adding to @insert
 158                  *
 159                  * Fist, drop rewrite_ptrs from @new:
 160                  */
 161                 i = 0;
 162                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
 163                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 164                             (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
 165                             !ptr->cached) {
 166                                 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr);
 167                                 rewrites_found |= 1U << i;
 168                         }
 169                         i++;
 170                 }
 171
 172                 if (m->data_opts.rewrite_ptrs &&
 173                     !rewrites_found &&
 174                     bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) {
 175                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:");
 176                         goto nowork;
 177                 }
 178
 179                 /*
 180                  * A replica that we just wrote might conflict with a replica
 181                  * that we want to keep, due to racing with another move:
 182                  */
 183 restart_drop_conflicting_replicas:
 184                 extent_for_each_ptr(extent_i_to_s(new), ptr)
 185                         if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
 186                             !ptr_c->cached) {
 187                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
 188                                 goto restart_drop_conflicting_replicas;
 189                         }
 190
 191                 if (!bkey_val_u64s(&new->k)) {
 192                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:");
 193                         goto nowork;
 194                 }
 195
 196                 /* Now, drop pointers that conflict with what we just wrote: */
 197                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
 198                         if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev)))
 199                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
 200
 201                 durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) +
 202                         bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i));
 203
 204                 /* Now, drop excess replicas: */
 205 restart_drop_extra_replicas:
 206
 207                 rcu_read_lock();
 208                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
 209                         unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
 210
 211                         if (!p.ptr.cached &&
 212                             durability - ptr_durability >= m->op.opts.data_replicas) {
 213                                 durability -= ptr_durability;
 214
 215                                 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr);
 216                                 goto restart_drop_extra_replicas;
 217                         }
 218                 }
 219                 rcu_read_unlock();
 220
 221                 /* Finally, add the pointers we just wrote: */
 222                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
 223                         bch2_extent_ptr_decoded_append(insert, &p);
 224
 225                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 226                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 227
 228                 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
 229                                                  &should_check_enospc,
 230                                                  &i_sectors_delta,
 231                                                  &disk_sectors_delta);
 232                 if (ret)
 233                         goto err;
 234
 235                 if (disk_sectors_delta > (s64) op->res.sectors) {
 236                         ret = bch2_disk_reservation_add(c, &op->res,
 237                                                 disk_sectors_delta - op->res.sectors,
 238                                                 !should_check_enospc
 239                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 240                         if (ret)
 241                                 goto out;
 242                 }
 243
 244                 next_pos = insert->k.p;
 245
 246                 /*
 247                  * Check for nonce offset inconsistency:
 248                  * This is debug code - we've been seeing this bug rarely, and
 249                  * it's been hard to reproduce, so this should give us some more
 250                  * information when it does occur:
 251                  */
 252                 struct printbuf err = PRINTBUF;
 253                 int invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), __btree_node_type(0, m->btree_id), 0, &err);
 254                 printbuf_exit(&err);
 255
 256                 if (invalid) {
 257                         struct printbuf buf = PRINTBUF;
 258
 259                         prt_str(&buf, "about to insert invalid key in data update path");
 260                         prt_str(&buf, "\nold: ");
 261                         bch2_bkey_val_to_text(&buf, c, old);
 262                         prt_str(&buf, "\nk:   ");
 263                         bch2_bkey_val_to_text(&buf, c, k);
 264                         prt_str(&buf, "\nnew: ");
 265                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 266
 267                         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 268                         printbuf_exit(&buf);
 269
 270                         bch2_fatal_error(c);
 271                         goto out;
 272                 }
 273
 274                 if (trace_data_update_enabled()) {
 275                         struct printbuf buf = PRINTBUF;
 276
 277                         prt_str(&buf, "\nold: ");
 278                         bch2_bkey_val_to_text(&buf, c, old);
 279                         prt_str(&buf, "\nk:   ");
 280                         bch2_bkey_val_to_text(&buf, c, k);
 281                         prt_str(&buf, "\nnew: ");
 282                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 283
 284                         trace_data_update(c, buf.buf);
 285                         printbuf_exit(&buf);
 286                 }
 287
 288                 ret =   bch2_insert_snapshot_whiteouts(trans, m->btree_id,
 289                                                 k.k->p, bkey_start_pos(&insert->k)) ?:
 290                         bch2_insert_snapshot_whiteouts(trans, m->btree_id,
 291                                                 k.k->p, insert->k.p) ?:
 292                         bch2_bkey_set_needs_rebalance(c, insert, &op->opts) ?:
 293                         bch2_trans_update(trans, &iter, insert,
 294                                 BTREE_UPDATE_internal_snapshot_node) ?:
 295                         bch2_trans_commit(trans, &op->res,
 296                                 NULL,
 297                                 BCH_TRANS_COMMIT_no_check_rw|
 298                                 BCH_TRANS_COMMIT_no_enospc|
 299                                 m->data_opts.btree_insert_flags);
 300                 if (!ret) {
 301                         bch2_btree_iter_set_pos(&iter, next_pos);
 302
 303                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 304                         trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i));
 305                 }
 306 err:
 307                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 308                         ret = 0;
 309                 if (ret)
 310                         break;
 311 next:
 312                 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
 313                         bch2_keylist_pop_front(keys);
 314                         if (bch2_keylist_empty(keys))
 315                                 goto out;
 316                 }
 317                 continue;
 318 nowork:
 319                 if (m->stats) {
 320                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 321                         atomic64_inc(&m->stats->keys_raced);
 322                         atomic64_add(k.k->p.offset - iter.pos.offset,
 323                                      &m->stats->sectors_raced);
 324                 }
 325
 326                 count_event(c, move_extent_fail);
 327
 328                 bch2_btree_iter_advance(&iter);
 329                 goto next;
 330         }
 331 out:
 332         bch2_trans_iter_exit(trans, &iter);
 333         bch2_bkey_buf_exit(&_insert, c);
 334         bch2_bkey_buf_exit(&_new, c);
 335         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 336         return ret;
 337 }
 338
 339 int bch2_data_update_index_update(struct bch_write_op *op)
 340 {
 341         return bch2_trans_run(op->c, __bch2_data_update_index_update(trans, op));
 342 }
 343
 344 void bch2_data_update_read_done(struct data_update *m,
 345                                 struct bch_extent_crc_unpacked crc)
 346 {
 347         /* write bio must own pages: */
 348         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 349
 350         m->op.crc = crc;
 351         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 352
 353         closure_call(&m->op.cl, bch2_write, NULL, NULL);
 354 }
 355
 356 void bch2_data_update_exit(struct data_update *update)
 357 {
 358         struct bch_fs *c = update->op.c;
 359         struct bkey_ptrs_c ptrs =
 360                 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
 361
 362         bkey_for_each_ptr(ptrs, ptr) {
 363                 struct bch_dev *ca = bch2_dev_have_ref(c, ptr->dev);
 364                 if (c->opts.nocow_enabled)
 365                         bch2_bucket_nocow_unlock(&c->nocow_locks,
 366                                                  PTR_BUCKET_POS(ca, ptr), 0);
 367                 bch2_dev_put(ca);
 368         }
 369
 370         bch2_bkey_buf_exit(&update->k, c);
 371         bch2_disk_reservation_put(c, &update->op.res);
 372         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 373 }
 374
 375 static void bch2_update_unwritten_extent(struct btree_trans *trans,
 376                                   struct data_update *update)
 377 {
 378         struct bch_fs *c = update->op.c;
 379         struct bio *bio = &update->op.wbio.bio;
 380         struct bkey_i_extent *e;
 381         struct write_point *wp;
 382         struct closure cl;
 383         struct btree_iter iter;
 384         struct bkey_s_c k;
 385         int ret;
 386
 387         closure_init_stack(&cl);
 388         bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
 389
 390         while (bio_sectors(bio)) {
 391                 unsigned sectors = bio_sectors(bio);
 392
 393                 bch2_trans_begin(trans);
 394
 395                 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
 396                                      BTREE_ITER_slots);
 397                 ret = lockrestart_do(trans, ({
 398                         k = bch2_btree_iter_peek_slot(&iter);
 399                         bkey_err(k);
 400                 }));
 401                 bch2_trans_iter_exit(trans, &iter);
 402
 403                 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
 404                         break;
 405
 406                 e = bkey_extent_init(update->op.insert_keys.top);
 407                 e->k.p = update->op.pos;
 408
 409                 ret = bch2_alloc_sectors_start_trans(trans,
 410                                 update->op.target,
 411                                 false,
 412                                 update->op.write_point,
 413                                 &update->op.devs_have,
 414                                 update->op.nr_replicas,
 415                                 update->op.nr_replicas,
 416                                 update->op.watermark,
 417                                 0, &cl, &wp);
 418                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
 419                         bch2_trans_unlock(trans);
 420                         closure_sync(&cl);
 421                         continue;
 422                 }
 423
 424                 bch_err_fn_ratelimited(c, ret);
 425
 426                 if (ret)
 427                         return;
 428
 429                 sectors = min(sectors, wp->sectors_free);
 430
 431                 bch2_key_resize(&e->k, sectors);
 432
 433                 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
 434                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 435                 bch2_alloc_sectors_done(c, wp);
 436
 437                 bio_advance(bio, sectors << 9);
 438                 update->op.pos.offset += sectors;
 439
 440                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 441                         ptr->unwritten = true;
 442                 bch2_keylist_push(&update->op.insert_keys);
 443
 444                 ret = __bch2_data_update_index_update(trans, &update->op);
 445
 446                 bch2_open_buckets_put(c, &update->op.open_buckets);
 447
 448                 if (ret)
 449                         break;
 450         }
 451
 452         if (closure_nr_remaining(&cl) != 1) {
 453                 bch2_trans_unlock(trans);
 454                 closure_sync(&cl);
 455         }
 456 }
 457
 458 int bch2_extent_drop_ptrs(struct btree_trans *trans,
 459                           struct btree_iter *iter,
 460                           struct bkey_s_c k,
 461                           struct data_update_opts data_opts)
 462 {
 463         struct bch_fs *c = trans->c;
 464         struct bkey_i *n;
 465         int ret;
 466
 467         n = bch2_bkey_make_mut_noupdate(trans, k);
 468         ret = PTR_ERR_OR_ZERO(n);
 469         if (ret)
 470                 return ret;
 471
 472         while (data_opts.kill_ptrs) {
 473                 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
 474
 475                 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
 476                 data_opts.kill_ptrs ^= 1U << drop;
 477         }
 478
 479         /*
 480          * If the new extent no longer has any pointers, bch2_extent_normalize()
 481          * will do the appropriate thing with it (turning it into a
 482          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 483          */
 484         bch2_extent_normalize(c, bkey_i_to_s(n));
 485
 486         /*
 487          * Since we're not inserting through an extent iterator
 488          * (BTREE_ITER_all_snapshots iterators aren't extent iterators),
 489          * we aren't using the extent overwrite path to delete, we're
 490          * just using the normal key deletion path:
 491          */
 492         if (bkey_deleted(&n->k) && !(iter->flags & BTREE_ITER_is_extents))
 493                 n->k.size = 0;
 494
 495         return bch2_trans_relock(trans) ?:
 496                 bch2_trans_update(trans, iter, n, BTREE_UPDATE_internal_snapshot_node) ?:
 497                 bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
 498 }
 499
 500 int bch2_data_update_init(struct btree_trans *trans,
 501                           struct btree_iter *iter,
 502                           struct moving_context *ctxt,
 503                           struct data_update *m,
 504                           struct write_point_specifier wp,
 505                           struct bch_io_opts io_opts,
 506                           struct data_update_opts data_opts,
 507                           enum btree_id btree_id,
 508                           struct bkey_s_c k)
 509 {
 510         struct bch_fs *c = trans->c;
 511         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 512         const union bch_extent_entry *entry;
 513         struct extent_ptr_decoded p;
 514         unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 515         unsigned ptrs_locked = 0;
 516         int ret = 0;
 517
 518         /*
 519          * fs is corrupt  we have a key for a snapshot node that doesn't exist,
 520          * and we have to check for this because we go rw before repairing the
 521          * snapshots table - just skip it, we can move it later.
 522          */
 523         if (unlikely(k.k->p.snapshot && !bch2_snapshot_equiv(c, k.k->p.snapshot)))
 524                 return -BCH_ERR_data_update_done;
 525
 526         bch2_bkey_buf_init(&m->k);
 527         bch2_bkey_buf_reassemble(&m->k, c, k);
 528         m->btree_id     = btree_id;
 529         m->data_opts    = data_opts;
 530         m->ctxt         = ctxt;
 531         m->stats        = ctxt ? ctxt->stats : NULL;
 532
 533         bch2_write_op_init(&m->op, c, io_opts);
 534         m->op.pos       = bkey_start_pos(k.k);
 535         m->op.version   = k.k->version;
 536         m->op.target    = data_opts.target;
 537         m->op.write_point = wp;
 538         m->op.nr_replicas = 0;
 539         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 540                 BCH_WRITE_PAGES_OWNED|
 541                 BCH_WRITE_DATA_ENCODED|
 542                 BCH_WRITE_MOVE|
 543                 m->data_opts.write_flags;
 544         m->op.compression_opt   = background_compression(io_opts);
 545         m->op.watermark         = m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
 546
 547         bkey_for_each_ptr(ptrs, ptr) {
 548                 if (!bch2_dev_tryget(c, ptr->dev)) {
 549                         bkey_for_each_ptr(ptrs, ptr2) {
 550                                 if (ptr2 == ptr)
 551                                         break;
 552                                 bch2_dev_put(bch2_dev_have_ref(c, ptr2->dev));
 553                         }
 554                         return -BCH_ERR_data_update_done;
 555                 }
 556         }
 557
 558         unsigned durability_have = 0, durability_removing = 0;
 559
 560         i = 0;
 561         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 562                 struct bch_dev *ca = bch2_dev_have_ref(c, p.ptr.dev);
 563                 struct bpos bucket = PTR_BUCKET_POS(ca, &p.ptr);
 564                 bool locked;
 565
 566                 rcu_read_lock();
 567                 if (((1U << i) & m->data_opts.rewrite_ptrs)) {
 568                         BUG_ON(p.ptr.cached);
 569
 570                         if (crc_is_compressed(p.crc))
 571                                 reserve_sectors += k.k->size;
 572
 573                         m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
 574                         durability_removing += bch2_extent_ptr_desired_durability(c, &p);
 575                 } else if (!p.ptr.cached &&
 576                            !((1U << i) & m->data_opts.kill_ptrs)) {
 577                         bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 578                         durability_have += bch2_extent_ptr_durability(c, &p);
 579                 }
 580                 rcu_read_unlock();
 581
 582                 /*
 583                  * op->csum_type is normally initialized from the fs/file's
 584                  * current options - but if an extent is encrypted, we require
 585                  * that it stays encrypted:
 586                  */
 587                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 588                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 589                         m->op.csum_type = p.crc.csum_type;
 590                 }
 591
 592                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 593                         m->op.incompressible = true;
 594
 595                 if (c->opts.nocow_enabled) {
 596                         if (ctxt) {
 597                                 move_ctxt_wait_event(ctxt,
 598                                                 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
 599                                                                           bucket, 0)) ||
 600                                                 list_empty(&ctxt->ios));
 601
 602                                 if (!locked)
 603                                         bch2_bucket_nocow_lock(&c->nocow_locks, bucket, 0);
 604                         } else {
 605                                 if (!bch2_bucket_nocow_trylock(&c->nocow_locks, bucket, 0)) {
 606                                         ret = -BCH_ERR_nocow_lock_blocked;
 607                                         goto err;
 608                                 }
 609                         }
 610                         ptrs_locked |= (1U << i);
 611                 }
 612
 613                 i++;
 614         }
 615
 616         unsigned durability_required = max(0, (int) (io_opts.data_replicas - durability_have));
 617
 618         /*
 619          * If current extent durability is less than io_opts.data_replicas,
 620          * we're not trying to rereplicate the extent up to data_replicas here -
 621          * unless extra_replicas was specified
 622          *
 623          * Increasing replication is an explicit operation triggered by
 624          * rereplicate, currently, so that users don't get an unexpected -ENOSPC
 625          */
 626         if (!(m->data_opts.write_flags & BCH_WRITE_CACHED) &&
 627             !durability_required) {
 628                 m->data_opts.kill_ptrs |= m->data_opts.rewrite_ptrs;
 629                 m->data_opts.rewrite_ptrs = 0;
 630                 /* if iter == NULL, it's just a promote */
 631                 if (iter)
 632                         ret = bch2_extent_drop_ptrs(trans, iter, k, m->data_opts);
 633                 goto done;
 634         }
 635
 636         m->op.nr_replicas = min(durability_removing, durability_required) +
 637                 m->data_opts.extra_replicas;
 638
 639         /*
 640          * If device(s) were set to durability=0 after data was written to them
 641          * we can end up with a duribilty=0 extent, and the normal algorithm
 642          * that tries not to increase durability doesn't work:
 643          */
 644         if (!(durability_have + durability_removing))
 645                 m->op.nr_replicas = max((unsigned) m->op.nr_replicas, 1);
 646
 647         m->op.nr_replicas_required = m->op.nr_replicas;
 648
 649         if (reserve_sectors) {
 650                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 651                                 m->data_opts.extra_replicas
 652                                 ? 0
 653                                 : BCH_DISK_RESERVATION_NOFAIL);
 654                 if (ret)
 655                         goto err;
 656         }
 657
 658         if (bkey_extent_is_unwritten(k)) {
 659                 bch2_update_unwritten_extent(trans, m);
 660                 goto done;
 661         }
 662
 663         return 0;
 664 err:
 665         i = 0;
 666         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 667                 struct bch_dev *ca = bch2_dev_have_ref(c, p.ptr.dev);
 668                 struct bpos bucket = PTR_BUCKET_POS(ca, &p.ptr);
 669                 if ((1U << i) & ptrs_locked)
 670                         bch2_bucket_nocow_unlock(&c->nocow_locks, bucket, 0);
 671                 bch2_dev_put(ca);
 672                 i++;
 673         }
 674
 675         bch2_bkey_buf_exit(&m->k, c);
 676         bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
 677         return ret;
 678 done:
 679         bch2_data_update_exit(m);
 680         return ret ?: -BCH_ERR_data_update_done;
 681 }
 682
 683 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 684 {
 685         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 686         unsigned i = 0;
 687
 688         bkey_for_each_ptr(ptrs, ptr) {
 689                 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
 690                         opts->kill_ptrs |= 1U << i;
 691                         opts->rewrite_ptrs ^= 1U << i;
 692                 }
 693
 694                 i++;
 695         }
 696 }