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 void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k)
  95 {
  96         if (trace_move_extent_finish_enabled()) {
  97                 struct printbuf buf = PRINTBUF;
  98
  99                 bch2_bkey_val_to_text(&buf, c, k);
 100                 trace_move_extent_finish(c, buf.buf);
 101                 printbuf_exit(&buf);
 102         }
 103 }
 104
 105 static void trace_move_extent_fail2(struct data_update *m,
 106                          struct bkey_s_c new,
 107                          struct bkey_s_c wrote,
 108                          struct bkey_i *insert,
 109                          const char *msg)
 110 {
 111         struct bch_fs *c = m->op.c;
 112         struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 113         const union bch_extent_entry *entry;
 114         struct bch_extent_ptr *ptr;
 115         struct extent_ptr_decoded p;
 116         struct printbuf buf = PRINTBUF;
 117         unsigned i, rewrites_found = 0;
 118
 119         if (!trace_move_extent_fail_enabled())
 120                 return;
 121
 122         prt_str(&buf, msg);
 123
 124         if (insert) {
 125                 i = 0;
 126                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
 127                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 128                             (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
 129                             !ptr->cached)
 130                                 rewrites_found |= 1U << i;
 131                         i++;
 132                 }
 133         }
 134
 135         prt_printf(&buf, "\nrewrite ptrs:   %u%u%u%u",
 136                    (m->data_opts.rewrite_ptrs & (1 << 0)) != 0,
 137                    (m->data_opts.rewrite_ptrs & (1 << 1)) != 0,
 138                    (m->data_opts.rewrite_ptrs & (1 << 2)) != 0,
 139                    (m->data_opts.rewrite_ptrs & (1 << 3)) != 0);
 140
 141         prt_printf(&buf, "\nrewrites found: %u%u%u%u",
 142                    (rewrites_found & (1 << 0)) != 0,
 143                    (rewrites_found & (1 << 1)) != 0,
 144                    (rewrites_found & (1 << 2)) != 0,
 145                    (rewrites_found & (1 << 3)) != 0);
 146
 147         prt_str(&buf, "\nold:    ");
 148         bch2_bkey_val_to_text(&buf, c, old);
 149
 150         prt_str(&buf, "\nnew:    ");
 151         bch2_bkey_val_to_text(&buf, c, new);
 152
 153         prt_str(&buf, "\nwrote:  ");
 154         bch2_bkey_val_to_text(&buf, c, wrote);
 155
 156         if (insert) {
 157                 prt_str(&buf, "\ninsert: ");
 158                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 159         }
 160
 161         trace_move_extent_fail(c, buf.buf);
 162         printbuf_exit(&buf);
 163 }
 164
 165 static int __bch2_data_update_index_update(struct btree_trans *trans,
 166                                            struct bch_write_op *op)
 167 {
 168         struct bch_fs *c = op->c;
 169         struct btree_iter iter;
 170         struct data_update *m =
 171                 container_of(op, struct data_update, op);
 172         struct keylist *keys = &op->insert_keys;
 173         struct bkey_buf _new, _insert;
 174         int ret = 0;
 175
 176         bch2_bkey_buf_init(&_new);
 177         bch2_bkey_buf_init(&_insert);
 178         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 179
 180         bch2_trans_iter_init(trans, &iter, m->btree_id,
 181                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 182                              BTREE_ITER_slots|BTREE_ITER_intent);
 183
 184         while (1) {
 185                 struct bkey_s_c k;
 186                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 187                 struct bkey_i *insert = NULL;
 188                 struct bkey_i_extent *new;
 189                 const union bch_extent_entry *entry_c;
 190                 union bch_extent_entry *entry;
 191                 struct extent_ptr_decoded p;
 192                 struct bch_extent_ptr *ptr;
 193                 const struct bch_extent_ptr *ptr_c;
 194                 struct bpos next_pos;
 195                 bool should_check_enospc;
 196                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 197                 unsigned rewrites_found = 0, durability, i;
 198
 199                 bch2_trans_begin(trans);
 200
 201                 k = bch2_btree_iter_peek_slot(&iter);
 202                 ret = bkey_err(k);
 203                 if (ret)
 204                         goto err;
 205
 206                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 207
 208                 if (!bch2_extents_match(k, old)) {
 209                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i),
 210                                                 NULL, "no match:");
 211                         goto nowork;
 212                 }
 213
 214                 bkey_reassemble(_insert.k, k);
 215                 insert = _insert.k;
 216
 217                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 218                 new = bkey_i_to_extent(_new.k);
 219                 bch2_cut_front(iter.pos, &new->k_i);
 220
 221                 bch2_cut_front(iter.pos,        insert);
 222                 bch2_cut_back(new->k.p,         insert);
 223                 bch2_cut_back(insert->k.p,      &new->k_i);
 224
 225                 /*
 226                  * @old: extent that we read from
 227                  * @insert: key that we're going to update, initialized from
 228                  * extent currently in btree - same as @old unless we raced with
 229                  * other updates
 230                  * @new: extent with new pointers that we'll be adding to @insert
 231                  *
 232                  * Fist, drop rewrite_ptrs from @new:
 233                  */
 234                 i = 0;
 235                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
 236                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 237                             (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
 238                             !ptr->cached) {
 239                                 bch2_extent_ptr_set_cached(c, &m->op.opts,
 240                                                            bkey_i_to_s(insert), ptr);
 241                                 rewrites_found |= 1U << i;
 242                         }
 243                         i++;
 244                 }
 245
 246                 if (m->data_opts.rewrite_ptrs &&
 247                     !rewrites_found &&
 248                     bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) {
 249                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:");
 250                         goto nowork;
 251                 }
 252
 253                 /*
 254                  * A replica that we just wrote might conflict with a replica
 255                  * that we want to keep, due to racing with another move:
 256                  */
 257 restart_drop_conflicting_replicas:
 258                 extent_for_each_ptr(extent_i_to_s(new), ptr)
 259                         if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
 260                             !ptr_c->cached) {
 261                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
 262                                 goto restart_drop_conflicting_replicas;
 263                         }
 264
 265                 if (!bkey_val_u64s(&new->k)) {
 266                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:");
 267                         goto nowork;
 268                 }
 269
 270                 /* Now, drop pointers that conflict with what we just wrote: */
 271                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
 272                         if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev)))
 273                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
 274
 275                 durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) +
 276                         bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i));
 277
 278                 /* Now, drop excess replicas: */
 279                 rcu_read_lock();
 280 restart_drop_extra_replicas:
 281                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
 282                         unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
 283
 284                         if (!p.ptr.cached &&
 285                             durability - ptr_durability >= m->op.opts.data_replicas) {
 286                                 durability -= ptr_durability;
 287
 288                                 bch2_extent_ptr_set_cached(c, &m->op.opts,
 289                                                            bkey_i_to_s(insert), &entry->ptr);
 290                                 goto restart_drop_extra_replicas;
 291                         }
 292                 }
 293                 rcu_read_unlock();
 294
 295                 /* Finally, add the pointers we just wrote: */
 296                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
 297                         bch2_extent_ptr_decoded_append(insert, &p);
 298
 299                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 300                 bch2_extent_normalize_by_opts(c, &m->op.opts, bkey_i_to_s(insert));
 301
 302                 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
 303                                                  &should_check_enospc,
 304                                                  &i_sectors_delta,
 305                                                  &disk_sectors_delta);
 306                 if (ret)
 307                         goto err;
 308
 309                 if (disk_sectors_delta > (s64) op->res.sectors) {
 310                         ret = bch2_disk_reservation_add(c, &op->res,
 311                                                 disk_sectors_delta - op->res.sectors,
 312                                                 !should_check_enospc
 313                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 314                         if (ret)
 315                                 goto out;
 316                 }
 317
 318                 next_pos = insert->k.p;
 319
 320                 /*
 321                  * Check for nonce offset inconsistency:
 322                  * This is debug code - we've been seeing this bug rarely, and
 323                  * it's been hard to reproduce, so this should give us some more
 324                  * information when it does occur:
 325                  */
 326                 int invalid = bch2_bkey_validate(c, bkey_i_to_s_c(insert), __btree_node_type(0, m->btree_id),
 327                                                  BCH_VALIDATE_commit);
 328                 if (invalid) {
 329                         struct printbuf buf = PRINTBUF;
 330
 331                         prt_str(&buf, "about to insert invalid key in data update path");
 332                         prt_str(&buf, "\nold: ");
 333                         bch2_bkey_val_to_text(&buf, c, old);
 334                         prt_str(&buf, "\nk:   ");
 335                         bch2_bkey_val_to_text(&buf, c, k);
 336                         prt_str(&buf, "\nnew: ");
 337                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 338
 339                         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 340                         printbuf_exit(&buf);
 341
 342                         bch2_fatal_error(c);
 343                         ret = -EIO;
 344                         goto out;
 345                 }
 346
 347                 if (trace_data_update_enabled()) {
 348                         struct printbuf buf = PRINTBUF;
 349
 350                         prt_str(&buf, "\nold: ");
 351                         bch2_bkey_val_to_text(&buf, c, old);
 352                         prt_str(&buf, "\nk:   ");
 353                         bch2_bkey_val_to_text(&buf, c, k);
 354                         prt_str(&buf, "\nnew: ");
 355                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 356
 357                         trace_data_update(c, buf.buf);
 358                         printbuf_exit(&buf);
 359                 }
 360
 361                 ret =   bch2_insert_snapshot_whiteouts(trans, m->btree_id,
 362                                                 k.k->p, bkey_start_pos(&insert->k)) ?:
 363                         bch2_insert_snapshot_whiteouts(trans, m->btree_id,
 364                                                 k.k->p, insert->k.p) ?:
 365                         bch2_bkey_set_needs_rebalance(c, insert, &op->opts) ?:
 366                         bch2_trans_update(trans, &iter, insert,
 367                                 BTREE_UPDATE_internal_snapshot_node) ?:
 368                         bch2_trans_commit(trans, &op->res,
 369                                 NULL,
 370                                 BCH_TRANS_COMMIT_no_check_rw|
 371                                 BCH_TRANS_COMMIT_no_enospc|
 372                                 m->data_opts.btree_insert_flags);
 373                 if (!ret) {
 374                         bch2_btree_iter_set_pos(&iter, next_pos);
 375
 376                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 377                         trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i));
 378                 }
 379 err:
 380                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 381                         ret = 0;
 382                 if (ret)
 383                         break;
 384 next:
 385                 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
 386                         bch2_keylist_pop_front(keys);
 387                         if (bch2_keylist_empty(keys))
 388                                 goto out;
 389                 }
 390                 continue;
 391 nowork:
 392                 if (m->stats) {
 393                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 394                         atomic64_inc(&m->stats->keys_raced);
 395                         atomic64_add(k.k->p.offset - iter.pos.offset,
 396                                      &m->stats->sectors_raced);
 397                 }
 398
 399                 count_event(c, move_extent_fail);
 400
 401                 bch2_btree_iter_advance(&iter);
 402                 goto next;
 403         }
 404 out:
 405         bch2_trans_iter_exit(trans, &iter);
 406         bch2_bkey_buf_exit(&_insert, c);
 407         bch2_bkey_buf_exit(&_new, c);
 408         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 409         return ret;
 410 }
 411
 412 int bch2_data_update_index_update(struct bch_write_op *op)
 413 {
 414         return bch2_trans_run(op->c, __bch2_data_update_index_update(trans, op));
 415 }
 416
 417 void bch2_data_update_read_done(struct data_update *m,
 418                                 struct bch_extent_crc_unpacked crc)
 419 {
 420         /* write bio must own pages: */
 421         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 422
 423         m->op.crc = crc;
 424         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 425
 426         closure_call(&m->op.cl, bch2_write, NULL, NULL);
 427 }
 428
 429 void bch2_data_update_exit(struct data_update *update)
 430 {
 431         struct bch_fs *c = update->op.c;
 432         struct bkey_s_c k = bkey_i_to_s_c(update->k.k);
 433
 434         if (c->opts.nocow_enabled)
 435                 bkey_nocow_unlock(c, k);
 436         bkey_put_dev_refs(c, k);
 437         bch2_bkey_buf_exit(&update->k, c);
 438         bch2_disk_reservation_put(c, &update->op.res);
 439         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 440 }
 441
 442 static void bch2_update_unwritten_extent(struct btree_trans *trans,
 443                                   struct data_update *update)
 444 {
 445         struct bch_fs *c = update->op.c;
 446         struct bio *bio = &update->op.wbio.bio;
 447         struct bkey_i_extent *e;
 448         struct write_point *wp;
 449         struct closure cl;
 450         struct btree_iter iter;
 451         struct bkey_s_c k;
 452         int ret;
 453
 454         closure_init_stack(&cl);
 455         bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
 456
 457         while (bio_sectors(bio)) {
 458                 unsigned sectors = bio_sectors(bio);
 459
 460                 bch2_trans_begin(trans);
 461
 462                 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
 463                                      BTREE_ITER_slots);
 464                 ret = lockrestart_do(trans, ({
 465                         k = bch2_btree_iter_peek_slot(&iter);
 466                         bkey_err(k);
 467                 }));
 468                 bch2_trans_iter_exit(trans, &iter);
 469
 470                 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
 471                         break;
 472
 473                 e = bkey_extent_init(update->op.insert_keys.top);
 474                 e->k.p = update->op.pos;
 475
 476                 ret = bch2_alloc_sectors_start_trans(trans,
 477                                 update->op.target,
 478                                 false,
 479                                 update->op.write_point,
 480                                 &update->op.devs_have,
 481                                 update->op.nr_replicas,
 482                                 update->op.nr_replicas,
 483                                 update->op.watermark,
 484                                 0, &cl, &wp);
 485                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
 486                         bch2_trans_unlock(trans);
 487                         closure_sync(&cl);
 488                         continue;
 489                 }
 490
 491                 bch_err_fn_ratelimited(c, ret);
 492
 493                 if (ret)
 494                         return;
 495
 496                 sectors = min(sectors, wp->sectors_free);
 497
 498                 bch2_key_resize(&e->k, sectors);
 499
 500                 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
 501                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 502                 bch2_alloc_sectors_done(c, wp);
 503
 504                 bio_advance(bio, sectors << 9);
 505                 update->op.pos.offset += sectors;
 506
 507                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 508                         ptr->unwritten = true;
 509                 bch2_keylist_push(&update->op.insert_keys);
 510
 511                 ret = __bch2_data_update_index_update(trans, &update->op);
 512
 513                 bch2_open_buckets_put(c, &update->op.open_buckets);
 514
 515                 if (ret)
 516                         break;
 517         }
 518
 519         if (closure_nr_remaining(&cl) != 1) {
 520                 bch2_trans_unlock(trans);
 521                 closure_sync(&cl);
 522         }
 523 }
 524
 525 void bch2_data_update_opts_to_text(struct printbuf *out, struct bch_fs *c,
 526                                    struct bch_io_opts *io_opts,
 527                                    struct data_update_opts *data_opts)
 528 {
 529         printbuf_tabstop_push(out, 20);
 530         prt_str(out, "rewrite ptrs:\t");
 531         bch2_prt_u64_base2(out, data_opts->rewrite_ptrs);
 532         prt_newline(out);
 533
 534         prt_str(out, "kill ptrs:\t");
 535         bch2_prt_u64_base2(out, data_opts->kill_ptrs);
 536         prt_newline(out);
 537
 538         prt_str(out, "target:\t");
 539         bch2_target_to_text(out, c, data_opts->target);
 540         prt_newline(out);
 541
 542         prt_str(out, "compression:\t");
 543         bch2_compression_opt_to_text(out, background_compression(*io_opts));
 544         prt_newline(out);
 545
 546         prt_str(out, "opts.replicas:\t");
 547         prt_u64(out, io_opts->data_replicas);
 548
 549         prt_str(out, "extra replicas:\t");
 550         prt_u64(out, data_opts->extra_replicas);
 551 }
 552
 553 void bch2_data_update_to_text(struct printbuf *out, struct data_update *m)
 554 {
 555         bch2_bkey_val_to_text(out, m->op.c, bkey_i_to_s_c(m->k.k));
 556         prt_newline(out);
 557         bch2_data_update_opts_to_text(out, m->op.c, &m->op.opts, &m->data_opts);
 558 }
 559
 560 int bch2_extent_drop_ptrs(struct btree_trans *trans,
 561                           struct btree_iter *iter,
 562                           struct bkey_s_c k,
 563                           struct bch_io_opts *io_opts,
 564                           struct data_update_opts *data_opts)
 565 {
 566         struct bch_fs *c = trans->c;
 567         struct bkey_i *n;
 568         int ret;
 569
 570         n = bch2_bkey_make_mut_noupdate(trans, k);
 571         ret = PTR_ERR_OR_ZERO(n);
 572         if (ret)
 573                 return ret;
 574
 575         while (data_opts->kill_ptrs) {
 576                 unsigned i = 0, drop = __fls(data_opts->kill_ptrs);
 577
 578                 bch2_bkey_drop_ptrs_noerror(bkey_i_to_s(n), ptr, i++ == drop);
 579                 data_opts->kill_ptrs ^= 1U << drop;
 580         }
 581
 582         /*
 583          * If the new extent no longer has any pointers, bch2_extent_normalize()
 584          * will do the appropriate thing with it (turning it into a
 585          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 586          */
 587         bch2_extent_normalize_by_opts(c, io_opts, bkey_i_to_s(n));
 588
 589         /*
 590          * Since we're not inserting through an extent iterator
 591          * (BTREE_ITER_all_snapshots iterators aren't extent iterators),
 592          * we aren't using the extent overwrite path to delete, we're
 593          * just using the normal key deletion path:
 594          */
 595         if (bkey_deleted(&n->k) && !(iter->flags & BTREE_ITER_is_extents))
 596                 n->k.size = 0;
 597
 598         return bch2_trans_relock(trans) ?:
 599                 bch2_trans_update(trans, iter, n, BTREE_UPDATE_internal_snapshot_node) ?:
 600                 bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
 601 }
 602
 603 int bch2_data_update_init(struct btree_trans *trans,
 604                           struct btree_iter *iter,
 605                           struct moving_context *ctxt,
 606                           struct data_update *m,
 607                           struct write_point_specifier wp,
 608                           struct bch_io_opts io_opts,
 609                           struct data_update_opts data_opts,
 610                           enum btree_id btree_id,
 611                           struct bkey_s_c k)
 612 {
 613         struct bch_fs *c = trans->c;
 614         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 615         const union bch_extent_entry *entry;
 616         struct extent_ptr_decoded p;
 617         unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 618         int ret = 0;
 619
 620         /*
 621          * fs is corrupt  we have a key for a snapshot node that doesn't exist,
 622          * and we have to check for this because we go rw before repairing the
 623          * snapshots table - just skip it, we can move it later.
 624          */
 625         if (unlikely(k.k->p.snapshot && !bch2_snapshot_equiv(c, k.k->p.snapshot)))
 626                 return -BCH_ERR_data_update_done;
 627
 628         if (!bkey_get_dev_refs(c, k))
 629                 return -BCH_ERR_data_update_done;
 630
 631         if (c->opts.nocow_enabled &&
 632             !bkey_nocow_lock(c, ctxt, k)) {
 633                 bkey_put_dev_refs(c, k);
 634                 return -BCH_ERR_nocow_lock_blocked;
 635         }
 636
 637         bch2_bkey_buf_init(&m->k);
 638         bch2_bkey_buf_reassemble(&m->k, c, k);
 639         m->btree_id     = btree_id;
 640         m->data_opts    = data_opts;
 641         m->ctxt         = ctxt;
 642         m->stats        = ctxt ? ctxt->stats : NULL;
 643
 644         bch2_write_op_init(&m->op, c, io_opts);
 645         m->op.pos       = bkey_start_pos(k.k);
 646         m->op.version   = k.k->bversion;
 647         m->op.target    = data_opts.target;
 648         m->op.write_point = wp;
 649         m->op.nr_replicas = 0;
 650         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 651                 BCH_WRITE_PAGES_OWNED|
 652                 BCH_WRITE_DATA_ENCODED|
 653                 BCH_WRITE_MOVE|
 654                 m->data_opts.write_flags;
 655         m->op.compression_opt   = background_compression(io_opts);
 656         m->op.watermark         = m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
 657
 658         unsigned durability_have = 0, durability_removing = 0;
 659
 660         i = 0;
 661         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 662                 if (!p.ptr.cached) {
 663                         rcu_read_lock();
 664                         if (BIT(i) & m->data_opts.rewrite_ptrs) {
 665                                 if (crc_is_compressed(p.crc))
 666                                         reserve_sectors += k.k->size;
 667
 668                                 m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
 669                                 durability_removing += bch2_extent_ptr_desired_durability(c, &p);
 670                         } else if (!(BIT(i) & m->data_opts.kill_ptrs)) {
 671                                 bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 672                                 durability_have += bch2_extent_ptr_durability(c, &p);
 673                         }
 674                         rcu_read_unlock();
 675                 }
 676
 677                 /*
 678                  * op->csum_type is normally initialized from the fs/file's
 679                  * current options - but if an extent is encrypted, we require
 680                  * that it stays encrypted:
 681                  */
 682                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 683                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 684                         m->op.csum_type = p.crc.csum_type;
 685                 }
 686
 687                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 688                         m->op.incompressible = true;
 689
 690                 i++;
 691         }
 692
 693         unsigned durability_required = max(0, (int) (io_opts.data_replicas - durability_have));
 694
 695         /*
 696          * If current extent durability is less than io_opts.data_replicas,
 697          * we're not trying to rereplicate the extent up to data_replicas here -
 698          * unless extra_replicas was specified
 699          *
 700          * Increasing replication is an explicit operation triggered by
 701          * rereplicate, currently, so that users don't get an unexpected -ENOSPC
 702          */
 703         m->op.nr_replicas = min(durability_removing, durability_required) +
 704                 m->data_opts.extra_replicas;
 705
 706         /*
 707          * If device(s) were set to durability=0 after data was written to them
 708          * we can end up with a duribilty=0 extent, and the normal algorithm
 709          * that tries not to increase durability doesn't work:
 710          */
 711         if (!(durability_have + durability_removing))
 712                 m->op.nr_replicas = max((unsigned) m->op.nr_replicas, 1);
 713
 714         m->op.nr_replicas_required = m->op.nr_replicas;
 715
 716         /*
 717          * It might turn out that we don't need any new replicas, if the
 718          * replicas or durability settings have been changed since the extent
 719          * was written:
 720          */
 721         if (!m->op.nr_replicas) {
 722                 m->data_opts.kill_ptrs |= m->data_opts.rewrite_ptrs;
 723                 m->data_opts.rewrite_ptrs = 0;
 724                 /* if iter == NULL, it's just a promote */
 725                 if (iter)
 726                         ret = bch2_extent_drop_ptrs(trans, iter, k, &io_opts, &m->data_opts);
 727                 goto out;
 728         }
 729
 730         if (reserve_sectors) {
 731                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 732                                 m->data_opts.extra_replicas
 733                                 ? 0
 734                                 : BCH_DISK_RESERVATION_NOFAIL);
 735                 if (ret)
 736                         goto out;
 737         }
 738
 739         if (bkey_extent_is_unwritten(k)) {
 740                 bch2_update_unwritten_extent(trans, m);
 741                 goto out;
 742         }
 743
 744         return 0;
 745 out:
 746         bch2_data_update_exit(m);
 747         return ret ?: -BCH_ERR_data_update_done;
 748 }
 749
 750 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 751 {
 752         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 753         unsigned i = 0;
 754
 755         bkey_for_each_ptr(ptrs, ptr) {
 756                 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
 757                         opts->kill_ptrs |= 1U << i;
 758                         opts->rewrite_ptrs ^= 1U << i;
 759                 }
 760
 761                 i++;
 762         }
 763 }