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 void bch2_bkey_mark_dev_cached(struct bkey_s k, unsigned dev)
  96 {
  97         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
  98         struct bch_extent_ptr *ptr;
  99
 100         bkey_for_each_ptr(ptrs, ptr)
 101                 if (ptr->dev == dev)
 102                         ptr->cached = true;
 103 }
 104
 105 static int __bch2_data_update_index_update(struct btree_trans *trans,
 106                                            struct bch_write_op *op)
 107 {
 108         struct bch_fs *c = op->c;
 109         struct btree_iter iter;
 110         struct data_update *m =
 111                 container_of(op, struct data_update, op);
 112         struct keylist *keys = &op->insert_keys;
 113         struct bkey_buf _new, _insert;
 114         int ret = 0;
 115
 116         bch2_bkey_buf_init(&_new);
 117         bch2_bkey_buf_init(&_insert);
 118         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 119
 120         bch2_trans_iter_init(trans, &iter, m->btree_id,
 121                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 122                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 123
 124         while (1) {
 125                 struct bkey_s_c k;
 126                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 127                 struct bkey_i *insert;
 128                 struct bkey_i_extent *new;
 129                 const union bch_extent_entry *entry;
 130                 struct extent_ptr_decoded p;
 131                 struct bpos next_pos;
 132                 bool did_work = false;
 133                 bool should_check_enospc;
 134                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 135                 unsigned i;
 136
 137                 bch2_trans_begin(trans);
 138
 139                 k = bch2_btree_iter_peek_slot(&iter);
 140                 ret = bkey_err(k);
 141                 if (ret)
 142                         goto err;
 143
 144                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 145
 146                 if (!bch2_extents_match(k, old))
 147                         goto nomatch;
 148
 149                 bkey_reassemble(_insert.k, k);
 150                 insert = _insert.k;
 151
 152                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 153                 new = bkey_i_to_extent(_new.k);
 154                 bch2_cut_front(iter.pos, &new->k_i);
 155
 156                 bch2_cut_front(iter.pos,        insert);
 157                 bch2_cut_back(new->k.p,         insert);
 158                 bch2_cut_back(insert->k.p,      &new->k_i);
 159
 160                 /*
 161                  * @old: extent that we read from
 162                  * @insert: key that we're going to update, initialized from
 163                  * extent currently in btree - same as @old unless we raced with
 164                  * other updates
 165                  * @new: extent with new pointers that we'll be adding to @insert
 166                  *
 167                  * Fist, drop rewrite_ptrs from @new:
 168                  */
 169                 i = 0;
 170                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
 171                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 172                             bch2_extent_has_ptr(old, p, bkey_i_to_s_c(insert))) {
 173                                 /*
 174                                  * If we're going to be adding a pointer to the
 175                                  * same device, we have to drop the old one -
 176                                  * otherwise, we can just mark it cached:
 177                                  */
 178                                 if (bch2_bkey_has_device(bkey_i_to_s_c(&new->k_i), p.ptr.dev))
 179                                         bch2_bkey_drop_device_noerror(bkey_i_to_s(insert), p.ptr.dev);
 180                                 else
 181                                         bch2_bkey_mark_dev_cached(bkey_i_to_s(insert), p.ptr.dev);
 182                         }
 183                         i++;
 184                 }
 185
 186
 187                 /* Add new ptrs: */
 188                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 189                         const struct bch_extent_ptr *existing_ptr =
 190                                 bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev);
 191
 192                         if (existing_ptr && existing_ptr->cached) {
 193                                 /*
 194                                  * We're replacing a cached pointer with a non
 195                                  * cached pointer:
 196                                  */
 197                                 bch2_bkey_drop_device_noerror(bkey_i_to_s(insert),
 198                                                               existing_ptr->dev);
 199                         } else if (existing_ptr) {
 200                                 /*
 201                                  * raced with another move op? extent already
 202                                  * has a pointer to the device we just wrote
 203                                  * data to
 204                                  */
 205                                 continue;
 206                         }
 207
 208                         bch2_extent_ptr_decoded_append(insert, &p);
 209                         did_work = true;
 210                 }
 211
 212                 if (!did_work)
 213                         goto nomatch;
 214
 215                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 216                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 217
 218                 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
 219                                                  &should_check_enospc,
 220                                                  &i_sectors_delta,
 221                                                  &disk_sectors_delta);
 222                 if (ret)
 223                         goto err;
 224
 225                 if (disk_sectors_delta > (s64) op->res.sectors) {
 226                         ret = bch2_disk_reservation_add(c, &op->res,
 227                                                 disk_sectors_delta - op->res.sectors,
 228                                                 !should_check_enospc
 229                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 230                         if (ret)
 231                                 goto out;
 232                 }
 233
 234                 next_pos = insert->k.p;
 235
 236                 if (!bkey_eq(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
 237                         ret = insert_snapshot_whiteouts(trans, m->btree_id, k.k->p,
 238                                                         bkey_start_pos(&insert->k));
 239                         if (ret)
 240                                 goto err;
 241                 }
 242
 243                 if (!bkey_eq(insert->k.p, k.k->p)) {
 244                         ret = insert_snapshot_whiteouts(trans, m->btree_id,
 245                                                         k.k->p, insert->k.p);
 246                         if (ret)
 247                                 goto err;
 248                 }
 249
 250                 ret   = bch2_trans_update(trans, &iter, insert,
 251                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 252                         bch2_trans_commit(trans, &op->res,
 253                                 NULL,
 254                                 BTREE_INSERT_NOFAIL|
 255                                 m->data_opts.btree_insert_flags);
 256                 if (!ret) {
 257                         bch2_btree_iter_set_pos(&iter, next_pos);
 258
 259                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 260                         trace_move_extent_finish(&new->k);
 261                 }
 262 err:
 263                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 264                         ret = 0;
 265                 if (ret)
 266                         break;
 267 next:
 268                 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
 269                         bch2_keylist_pop_front(keys);
 270                         if (bch2_keylist_empty(keys))
 271                                 goto out;
 272                 }
 273                 continue;
 274 nomatch:
 275                 if (m->ctxt) {
 276                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 277                         atomic64_inc(&m->ctxt->stats->keys_raced);
 278                         atomic64_add(k.k->p.offset - iter.pos.offset,
 279                                      &m->ctxt->stats->sectors_raced);
 280                 }
 281
 282                 this_cpu_add(c->counters[BCH_COUNTER_move_extent_fail], new->k.size);
 283                 trace_move_extent_fail(&new->k);
 284
 285                 bch2_btree_iter_advance(&iter);
 286                 goto next;
 287         }
 288 out:
 289         bch2_trans_iter_exit(trans, &iter);
 290         bch2_bkey_buf_exit(&_insert, c);
 291         bch2_bkey_buf_exit(&_new, c);
 292         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 293         return ret;
 294 }
 295
 296 int bch2_data_update_index_update(struct bch_write_op *op)
 297 {
 298         struct bch_fs *c = op->c;
 299         struct btree_trans trans;
 300         int ret;
 301
 302         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 303         ret = __bch2_data_update_index_update(&trans, op);
 304         bch2_trans_exit(&trans);
 305
 306         return ret;
 307 }
 308
 309 void bch2_data_update_read_done(struct data_update *m,
 310                                 struct bch_extent_crc_unpacked crc)
 311 {
 312         /* write bio must own pages: */
 313         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 314
 315         m->op.crc = crc;
 316         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 317
 318         closure_call(&m->op.cl, bch2_write, NULL, NULL);
 319 }
 320
 321 void bch2_data_update_exit(struct data_update *update)
 322 {
 323         struct bch_fs *c = update->op.c;
 324         struct bkey_ptrs_c ptrs =
 325                 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
 326         const struct bch_extent_ptr *ptr;
 327
 328         bkey_for_each_ptr(ptrs, ptr) {
 329                 bch2_bucket_nocow_unlock(&c->nocow_locks,
 330                                          PTR_BUCKET_POS(c, ptr), 0);
 331                 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 332         }
 333
 334         bch2_bkey_buf_exit(&update->k, c);
 335         bch2_disk_reservation_put(c, &update->op.res);
 336         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 337 }
 338
 339 void bch2_update_unwritten_extent(struct btree_trans *trans,
 340                                   struct data_update *update)
 341 {
 342         struct bch_fs *c = update->op.c;
 343         struct bio *bio = &update->op.wbio.bio;
 344         struct bkey_i_extent *e;
 345         struct write_point *wp;
 346         struct bch_extent_ptr *ptr;
 347         struct closure cl;
 348         struct btree_iter iter;
 349         struct bkey_s_c k;
 350         int ret;
 351
 352         closure_init_stack(&cl);
 353         bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
 354
 355         while (bio_sectors(bio)) {
 356                 unsigned sectors = bio_sectors(bio);
 357
 358                 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
 359                                      BTREE_ITER_SLOTS);
 360                 ret = lockrestart_do(trans, ({
 361                         k = bch2_btree_iter_peek_slot(&iter);
 362                         bkey_err(k);
 363                 }));
 364                 bch2_trans_iter_exit(trans, &iter);
 365
 366                 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
 367                         break;
 368
 369                 e = bkey_extent_init(update->op.insert_keys.top);
 370                 e->k.p = update->op.pos;
 371
 372                 ret = bch2_alloc_sectors_start_trans(trans,
 373                                 update->op.target,
 374                                 false,
 375                                 update->op.write_point,
 376                                 &update->op.devs_have,
 377                                 update->op.nr_replicas,
 378                                 update->op.nr_replicas,
 379                                 update->op.alloc_reserve,
 380                                 0, &cl, &wp);
 381                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
 382                         bch2_trans_unlock(trans);
 383                         closure_sync(&cl);
 384                         continue;
 385                 }
 386
 387                 if (ret)
 388                         return;
 389
 390                 sectors = min(sectors, wp->sectors_free);
 391
 392                 bch2_key_resize(&e->k, sectors);
 393
 394                 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
 395                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 396                 bch2_alloc_sectors_done(c, wp);
 397
 398                 bio_advance(bio, sectors << 9);
 399                 update->op.pos.offset += sectors;
 400
 401                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 402                         ptr->unwritten = true;
 403                 bch2_keylist_push(&update->op.insert_keys);
 404
 405                 ret = __bch2_data_update_index_update(trans, &update->op);
 406
 407                 bch2_open_buckets_put(c, &update->op.open_buckets);
 408
 409                 if (ret)
 410                         break;
 411         }
 412
 413         if ((atomic_read(&cl.remaining) & CLOSURE_REMAINING_MASK) != 1) {
 414                 bch2_trans_unlock(trans);
 415                 closure_sync(&cl);
 416         }
 417 }
 418
 419 int bch2_data_update_init(struct btree_trans *trans,
 420                           struct moving_context *ctxt,
 421                           struct data_update *m,
 422                           struct write_point_specifier wp,
 423                           struct bch_io_opts io_opts,
 424                           struct data_update_opts data_opts,
 425                           enum btree_id btree_id,
 426                           struct bkey_s_c k)
 427 {
 428         struct bch_fs *c = trans->c;
 429         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 430         const union bch_extent_entry *entry;
 431         struct extent_ptr_decoded p;
 432         const struct bch_extent_ptr *ptr;
 433         unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 434         unsigned int ptrs_locked = 0;
 435         int ret;
 436
 437         bch2_bkey_buf_init(&m->k);
 438         bch2_bkey_buf_reassemble(&m->k, c, k);
 439         m->btree_id     = btree_id;
 440         m->data_opts    = data_opts;
 441
 442         bch2_write_op_init(&m->op, c, io_opts);
 443         m->op.pos       = bkey_start_pos(k.k);
 444         m->op.version   = k.k->version;
 445         m->op.target    = data_opts.target;
 446         m->op.write_point = wp;
 447         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 448                 BCH_WRITE_PAGES_OWNED|
 449                 BCH_WRITE_DATA_ENCODED|
 450                 BCH_WRITE_MOVE|
 451                 m->data_opts.write_flags;
 452         m->op.compression_type =
 453                 bch2_compression_opt_to_type[io_opts.background_compression ?:
 454                                              io_opts.compression];
 455         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 456                 m->op.alloc_reserve = RESERVE_movinggc;
 457
 458         bkey_for_each_ptr(ptrs, ptr)
 459                 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 460
 461         i = 0;
 462         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 463                 bool locked;
 464
 465                 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 466                     p.ptr.cached)
 467                         BUG();
 468
 469                 if (!((1U << i) & m->data_opts.rewrite_ptrs) &&
 470                     !p.ptr.cached)
 471                         bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 472
 473                 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 474                     crc_is_compressed(p.crc))
 475                         reserve_sectors += k.k->size;
 476
 477                 /*
 478                  * op->csum_type is normally initialized from the fs/file's
 479                  * current options - but if an extent is encrypted, we require
 480                  * that it stays encrypted:
 481                  */
 482                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 483                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 484                         m->op.csum_type = p.crc.csum_type;
 485                 }
 486
 487                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 488                         m->op.incompressible = true;
 489
 490                 if (ctxt) {
 491                         move_ctxt_wait_event(ctxt, trans,
 492                                         (locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
 493                                                                   PTR_BUCKET_POS(c, &p.ptr), 0)) ||
 494                                         !atomic_read(&ctxt->read_sectors));
 495
 496                         if (!locked)
 497                                 bch2_bucket_nocow_lock(&c->nocow_locks,
 498                                                        PTR_BUCKET_POS(c, &p.ptr), 0);
 499                 } else {
 500                         if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
 501                                                        PTR_BUCKET_POS(c, &p.ptr), 0)) {
 502                                 ret = -BCH_ERR_nocow_lock_blocked;
 503                                 goto err;
 504                         }
 505                 }
 506                 ptrs_locked |= (1U << i);
 507                 i++;
 508         }
 509
 510         if (reserve_sectors) {
 511                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 512                                 m->data_opts.extra_replicas
 513                                 ? 0
 514                                 : BCH_DISK_RESERVATION_NOFAIL);
 515                 if (ret)
 516                         goto err;
 517         }
 518
 519         m->op.nr_replicas = m->op.nr_replicas_required =
 520                 hweight32(m->data_opts.rewrite_ptrs) + m->data_opts.extra_replicas;
 521
 522         BUG_ON(!m->op.nr_replicas);
 523
 524         /* Special handling required: */
 525         if (bkey_extent_is_unwritten(k))
 526                 return -BCH_ERR_unwritten_extent_update;
 527         return 0;
 528 err:
 529         i = 0;
 530         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 531                 if ((1U << i) & ptrs_locked)
 532                         bch2_bucket_nocow_unlock(&c->nocow_locks,
 533                                                  PTR_BUCKET_POS(c, &p.ptr), 0);
 534                 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
 535                 i++;
 536         }
 537
 538         bch2_bkey_buf_exit(&m->k, c);
 539         bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
 540         return ret;
 541 }
 542
 543 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 544 {
 545         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 546         const struct bch_extent_ptr *ptr;
 547         unsigned i = 0;
 548
 549         bkey_for_each_ptr(ptrs, ptr) {
 550                 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
 551                         opts->kill_ptrs |= 1U << i;
 552                         opts->rewrite_ptrs ^= 1U << i;
 553                 }
 554
 555                 i++;
 556         }
 557 }