git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/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 "extents.h"
  11 #include "io.h"
  12 #include "keylist.h"
  13 #include "move.h"
  14 #include "nocow_locking.h"
  15 #include "subvolume.h"
  16
  17 #include <trace/events/bcachefs.h>
  18
  19 static int insert_snapshot_whiteouts(struct btree_trans *trans,
  20                                      enum btree_id id,
  21                                      struct bpos old_pos,
  22                                      struct bpos new_pos)
  23 {
  24         struct bch_fs *c = trans->c;
  25         struct btree_iter iter, iter2;
  26         struct bkey_s_c k, k2;
  27         snapshot_id_list s;
  28         struct bkey_i *update;
  29         int ret;
  30
  31         if (!btree_type_has_snapshots(id))
  32                 return 0;
  33
  34         darray_init(&s);
  35
  36         if (!bch2_snapshot_has_children(c, old_pos.snapshot))
  37                 return 0;
  38
  39         bch2_trans_iter_init(trans, &iter, id, old_pos,
  40                              BTREE_ITER_NOT_EXTENTS|
  41                              BTREE_ITER_ALL_SNAPSHOTS);
  42         while (1) {
  43                 k = bch2_btree_iter_prev(&iter);
  44                 ret = bkey_err(k);
  45                 if (ret)
  46                         break;
  47
  48                 if (!k.k)
  49                         break;
  50
  51                 if (!bkey_eq(old_pos, k.k->p))
  52                         break;
  53
  54                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot) &&
  55                     !snapshot_list_has_ancestor(c, &s, k.k->p.snapshot)) {
  56                         struct bpos whiteout_pos = new_pos;
  57
  58                         whiteout_pos.snapshot = k.k->p.snapshot;
  59
  60                         bch2_trans_iter_init(trans, &iter2, id, whiteout_pos,
  61                                              BTREE_ITER_NOT_EXTENTS|
  62                                              BTREE_ITER_INTENT);
  63                         k2 = bch2_btree_iter_peek_slot(&iter2);
  64                         ret = bkey_err(k2);
  65
  66                         if (!ret && k2.k->type == KEY_TYPE_deleted) {
  67                                 update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
  68                                 ret = PTR_ERR_OR_ZERO(update);
  69                                 if (ret)
  70                                         break;
  71
  72                                 bkey_init(&update->k);
  73                                 update->k.p             = whiteout_pos;
  74                                 update->k.type          = KEY_TYPE_whiteout;
  75
  76                                 ret = bch2_trans_update(trans, &iter2, update,
  77                                                         BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
  78                         }
  79                         bch2_trans_iter_exit(trans, &iter2);
  80
  81                         if (ret)
  82                                 break;
  83
  84                         ret = snapshot_list_add(c, &s, k.k->p.snapshot);
  85                         if (ret)
  86                                 break;
  87                 }
  88         }
  89         bch2_trans_iter_exit(trans, &iter);
  90         darray_exit(&s);
  91
  92         return ret;
  93 }
  94
  95 static int __bch2_data_update_index_update(struct btree_trans *trans,
  96                                            struct bch_write_op *op)
  97 {
  98         struct bch_fs *c = op->c;
  99         struct btree_iter iter;
 100         struct data_update *m =
 101                 container_of(op, struct data_update, op);
 102         struct keylist *keys = &op->insert_keys;
 103         struct bkey_buf _new, _insert;
 104         int ret = 0;
 105
 106         bch2_bkey_buf_init(&_new);
 107         bch2_bkey_buf_init(&_insert);
 108         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 109
 110         bch2_trans_iter_init(trans, &iter, m->btree_id,
 111                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 112                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 113
 114         while (1) {
 115                 struct bkey_s_c k;
 116                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 117                 struct bkey_i *insert = NULL;
 118                 struct bkey_i_extent *new;
 119                 const union bch_extent_entry *entry_c;
 120                 union bch_extent_entry *entry;
 121                 struct extent_ptr_decoded p;
 122                 struct bch_extent_ptr *ptr;
 123                 const struct bch_extent_ptr *ptr_c;
 124                 struct bpos next_pos;
 125                 bool should_check_enospc;
 126                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 127                 unsigned rewrites_found = 0, durability, i;
 128
 129                 bch2_trans_begin(trans);
 130
 131                 k = bch2_btree_iter_peek_slot(&iter);
 132                 ret = bkey_err(k);
 133                 if (ret)
 134                         goto err;
 135
 136                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 137
 138                 if (!bch2_extents_match(k, old))
 139                         goto nowork;
 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_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
 167                                 /*
 168                                  * See comment below:
 169                                 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr);
 170                                 */
 171                                 rewrites_found |= 1U << i;
 172                         }
 173                         i++;
 174                 }
 175
 176                 if (m->data_opts.rewrite_ptrs &&
 177                     !rewrites_found &&
 178                     bch2_bkey_durability(c, k) >= m->op.opts.data_replicas)
 179                         goto nowork;
 180
 181                 /*
 182                  * A replica that we just wrote might conflict with a replica
 183                  * that we want to keep, due to racing with another move:
 184                  */
 185 restart_drop_conflicting_replicas:
 186                 extent_for_each_ptr(extent_i_to_s(new), ptr)
 187                         if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
 188                             !ptr_c->cached) {
 189                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
 190                                 goto restart_drop_conflicting_replicas;
 191                         }
 192
 193                 if (!bkey_val_u64s(&new->k))
 194                         goto nowork;
 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                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
 207                         unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
 208
 209                         if (!p.ptr.cached &&
 210                             durability - ptr_durability >= m->op.opts.data_replicas) {
 211                                 durability -= ptr_durability;
 212                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), &entry->ptr);
 213                                 /*
 214                                  * Currently, we're dropping unneeded replicas
 215                                  * instead of marking them as cached, since
 216                                  * cached data in stripe buckets prevents them
 217                                  * from being reused:
 218                                 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr);
 219                                 */
 220                                 goto restart_drop_extra_replicas;
 221                         }
 222                 }
 223
 224                 /* Finally, add the pointers we just wrote: */
 225                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
 226                         bch2_extent_ptr_decoded_append(insert, &p);
 227
 228                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 229                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 230
 231                 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
 232                                                  &should_check_enospc,
 233                                                  &i_sectors_delta,
 234                                                  &disk_sectors_delta);
 235                 if (ret)
 236                         goto err;
 237
 238                 if (disk_sectors_delta > (s64) op->res.sectors) {
 239                         ret = bch2_disk_reservation_add(c, &op->res,
 240                                                 disk_sectors_delta - op->res.sectors,
 241                                                 !should_check_enospc
 242                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 243                         if (ret)
 244                                 goto out;
 245                 }
 246
 247                 next_pos = insert->k.p;
 248
 249                 if (!bkey_eq(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
 250                         ret = insert_snapshot_whiteouts(trans, m->btree_id, k.k->p,
 251                                                         bkey_start_pos(&insert->k));
 252                         if (ret)
 253                                 goto err;
 254                 }
 255
 256                 if (!bkey_eq(insert->k.p, k.k->p)) {
 257                         ret = insert_snapshot_whiteouts(trans, m->btree_id,
 258                                                         k.k->p, insert->k.p);
 259                         if (ret)
 260                                 goto err;
 261                 }
 262
 263                 ret   = bch2_trans_update(trans, &iter, insert,
 264                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 265                         bch2_trans_commit(trans, &op->res,
 266                                 NULL,
 267                                 BTREE_INSERT_NOCHECK_RW|
 268                                 BTREE_INSERT_NOFAIL|
 269                                 m->data_opts.btree_insert_flags);
 270                 if (!ret) {
 271                         bch2_btree_iter_set_pos(&iter, next_pos);
 272
 273                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 274                         trace_move_extent_finish(&new->k);
 275                 }
 276 err:
 277                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 278                         ret = 0;
 279                 if (ret)
 280                         break;
 281 next:
 282                 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
 283                         bch2_keylist_pop_front(keys);
 284                         if (bch2_keylist_empty(keys))
 285                                 goto out;
 286                 }
 287                 continue;
 288 nowork:
 289                 if (m->ctxt && m->ctxt->stats) {
 290                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 291                         atomic64_inc(&m->ctxt->stats->keys_raced);
 292                         atomic64_add(k.k->p.offset - iter.pos.offset,
 293                                      &m->ctxt->stats->sectors_raced);
 294                 }
 295
 296                 this_cpu_add(c->counters[BCH_COUNTER_move_extent_fail], new->k.size);
 297                 trace_move_extent_fail(&new->k);
 298
 299                 bch2_btree_iter_advance(&iter);
 300                 goto next;
 301         }
 302 out:
 303         bch2_trans_iter_exit(trans, &iter);
 304         bch2_bkey_buf_exit(&_insert, c);
 305         bch2_bkey_buf_exit(&_new, c);
 306         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 307         return ret;
 308 }
 309
 310 int bch2_data_update_index_update(struct bch_write_op *op)
 311 {
 312         return bch2_trans_run(op->c, __bch2_data_update_index_update(&trans, op));
 313 }
 314
 315 void bch2_data_update_read_done(struct data_update *m,
 316                                 struct bch_extent_crc_unpacked crc)
 317 {
 318         /* write bio must own pages: */
 319         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 320
 321         m->op.crc = crc;
 322         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 323
 324         closure_call(&m->op.cl, bch2_write, NULL, NULL);
 325 }
 326
 327 void bch2_data_update_exit(struct data_update *update)
 328 {
 329         struct bch_fs *c = update->op.c;
 330         struct bkey_ptrs_c ptrs =
 331                 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
 332         const struct bch_extent_ptr *ptr;
 333
 334         bkey_for_each_ptr(ptrs, ptr) {
 335                 if (c->opts.nocow_enabled)
 336                         bch2_bucket_nocow_unlock(&c->nocow_locks,
 337                                                  PTR_BUCKET_POS(c, ptr), 0);
 338                 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 339         }
 340
 341         bch2_bkey_buf_exit(&update->k, c);
 342         bch2_disk_reservation_put(c, &update->op.res);
 343         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 344 }
 345
 346 void bch2_update_unwritten_extent(struct btree_trans *trans,
 347                                   struct data_update *update)
 348 {
 349         struct bch_fs *c = update->op.c;
 350         struct bio *bio = &update->op.wbio.bio;
 351         struct bkey_i_extent *e;
 352         struct write_point *wp;
 353         struct bch_extent_ptr *ptr;
 354         struct closure cl;
 355         struct btree_iter iter;
 356         struct bkey_s_c k;
 357         int ret;
 358
 359         closure_init_stack(&cl);
 360         bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
 361
 362         while (bio_sectors(bio)) {
 363                 unsigned sectors = bio_sectors(bio);
 364
 365                 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
 366                                      BTREE_ITER_SLOTS);
 367                 ret = lockrestart_do(trans, ({
 368                         k = bch2_btree_iter_peek_slot(&iter);
 369                         bkey_err(k);
 370                 }));
 371                 bch2_trans_iter_exit(trans, &iter);
 372
 373                 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
 374                         break;
 375
 376                 e = bkey_extent_init(update->op.insert_keys.top);
 377                 e->k.p = update->op.pos;
 378
 379                 ret = bch2_alloc_sectors_start_trans(trans,
 380                                 update->op.target,
 381                                 false,
 382                                 update->op.write_point,
 383                                 &update->op.devs_have,
 384                                 update->op.nr_replicas,
 385                                 update->op.nr_replicas,
 386                                 update->op.alloc_reserve,
 387                                 0, &cl, &wp);
 388                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
 389                         bch2_trans_unlock(trans);
 390                         closure_sync(&cl);
 391                         continue;
 392                 }
 393
 394                 if (ret)
 395                         return;
 396
 397                 sectors = min(sectors, wp->sectors_free);
 398
 399                 bch2_key_resize(&e->k, sectors);
 400
 401                 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
 402                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 403                 bch2_alloc_sectors_done(c, wp);
 404
 405                 bio_advance(bio, sectors << 9);
 406                 update->op.pos.offset += sectors;
 407
 408                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 409                         ptr->unwritten = true;
 410                 bch2_keylist_push(&update->op.insert_keys);
 411
 412                 ret = __bch2_data_update_index_update(trans, &update->op);
 413
 414                 bch2_open_buckets_put(c, &update->op.open_buckets);
 415
 416                 if (ret)
 417                         break;
 418         }
 419
 420         if ((atomic_read(&cl.remaining) & CLOSURE_REMAINING_MASK) != 1) {
 421                 bch2_trans_unlock(trans);
 422                 closure_sync(&cl);
 423         }
 424 }
 425
 426 int bch2_data_update_init(struct btree_trans *trans,
 427                           struct moving_context *ctxt,
 428                           struct data_update *m,
 429                           struct write_point_specifier wp,
 430                           struct bch_io_opts io_opts,
 431                           struct data_update_opts data_opts,
 432                           enum btree_id btree_id,
 433                           struct bkey_s_c k)
 434 {
 435         struct bch_fs *c = trans->c;
 436         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 437         const union bch_extent_entry *entry;
 438         struct extent_ptr_decoded p;
 439         const struct bch_extent_ptr *ptr;
 440         unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 441         unsigned ptrs_locked = 0;
 442         int ret;
 443
 444         bch2_bkey_buf_init(&m->k);
 445         bch2_bkey_buf_reassemble(&m->k, c, k);
 446         m->btree_id     = btree_id;
 447         m->data_opts    = data_opts;
 448
 449         bch2_write_op_init(&m->op, c, io_opts);
 450         m->op.pos       = bkey_start_pos(k.k);
 451         m->op.version   = k.k->version;
 452         m->op.target    = data_opts.target;
 453         m->op.write_point = wp;
 454         m->op.nr_replicas = 0;
 455         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 456                 BCH_WRITE_PAGES_OWNED|
 457                 BCH_WRITE_DATA_ENCODED|
 458                 BCH_WRITE_MOVE|
 459                 m->data_opts.write_flags;
 460         m->op.compression_type =
 461                 bch2_compression_opt_to_type[io_opts.background_compression ?:
 462                                              io_opts.compression];
 463         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 464                 m->op.alloc_reserve = RESERVE_movinggc;
 465
 466         bkey_for_each_ptr(ptrs, ptr)
 467                 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 468
 469         i = 0;
 470         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 471                 bool locked;
 472
 473                 if (((1U << i) & m->data_opts.rewrite_ptrs)) {
 474                         BUG_ON(p.ptr.cached);
 475
 476                         if (crc_is_compressed(p.crc))
 477                                 reserve_sectors += k.k->size;
 478
 479                         m->op.nr_replicas += bch2_extent_ptr_durability(c, &p);
 480                 } else if (!p.ptr.cached) {
 481                         bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 482                 }
 483
 484                 /*
 485                  * op->csum_type is normally initialized from the fs/file's
 486                  * current options - but if an extent is encrypted, we require
 487                  * that it stays encrypted:
 488                  */
 489                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 490                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 491                         m->op.csum_type = p.crc.csum_type;
 492                 }
 493
 494                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 495                         m->op.incompressible = true;
 496
 497                 if (c->opts.nocow_enabled) {
 498                         if (ctxt) {
 499                                 move_ctxt_wait_event(ctxt, trans,
 500                                                 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
 501                                                                           PTR_BUCKET_POS(c, &p.ptr), 0)) ||
 502                                                 !atomic_read(&ctxt->read_sectors));
 503
 504                                 if (!locked)
 505                                         bch2_bucket_nocow_lock(&c->nocow_locks,
 506                                                                PTR_BUCKET_POS(c, &p.ptr), 0);
 507                         } else {
 508                                 if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
 509                                                                PTR_BUCKET_POS(c, &p.ptr), 0)) {
 510                                         ret = -BCH_ERR_nocow_lock_blocked;
 511                                         goto err;
 512                                 }
 513                         }
 514                         ptrs_locked |= (1U << i);
 515                 }
 516
 517                 i++;
 518         }
 519
 520         if (reserve_sectors) {
 521                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 522                                 m->data_opts.extra_replicas
 523                                 ? 0
 524                                 : BCH_DISK_RESERVATION_NOFAIL);
 525                 if (ret)
 526                         goto err;
 527         }
 528
 529         m->op.nr_replicas += m->data_opts.extra_replicas;
 530         m->op.nr_replicas_required = m->op.nr_replicas;
 531
 532         BUG_ON(!m->op.nr_replicas);
 533
 534         /* Special handling required: */
 535         if (bkey_extent_is_unwritten(k))
 536                 return -BCH_ERR_unwritten_extent_update;
 537         return 0;
 538 err:
 539         i = 0;
 540         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 541                 if ((1U << i) & ptrs_locked)
 542                         bch2_bucket_nocow_unlock(&c->nocow_locks,
 543                                                  PTR_BUCKET_POS(c, &p.ptr), 0);
 544                 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
 545                 i++;
 546         }
 547
 548         bch2_bkey_buf_exit(&m->k, c);
 549         bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
 550         return ret;
 551 }
 552
 553 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 554 {
 555         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 556         const struct bch_extent_ptr *ptr;
 557         unsigned i = 0;
 558
 559         bkey_for_each_ptr(ptrs, ptr) {
 560                 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
 561                         opts->kill_ptrs |= 1U << i;
 562                         opts->rewrite_ptrs ^= 1U << i;
 563                 }
 564
 565                 i++;
 566         }
 567 }