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                         bch2_extent_ptr_set_cached(k, ptr);
 103                         return;
 104                 }
 105 }
 106
 107 static int __bch2_data_update_index_update(struct btree_trans *trans,
 108                                            struct bch_write_op *op)
 109 {
 110         struct bch_fs *c = op->c;
 111         struct btree_iter iter;
 112         struct data_update *m =
 113                 container_of(op, struct data_update, op);
 114         struct keylist *keys = &op->insert_keys;
 115         struct bkey_buf _new, _insert;
 116         int ret = 0;
 117
 118         bch2_bkey_buf_init(&_new);
 119         bch2_bkey_buf_init(&_insert);
 120         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 121
 122         bch2_trans_iter_init(trans, &iter, m->btree_id,
 123                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 124                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 125
 126         while (1) {
 127                 struct bkey_s_c k;
 128                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 129                 struct bkey_i *insert;
 130                 struct bkey_i_extent *new;
 131                 const union bch_extent_entry *entry;
 132                 struct extent_ptr_decoded p;
 133                 struct bpos next_pos;
 134                 bool did_work = false;
 135                 bool should_check_enospc;
 136                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 137                 unsigned i;
 138
 139                 bch2_trans_begin(trans);
 140
 141                 k = bch2_btree_iter_peek_slot(&iter);
 142                 ret = bkey_err(k);
 143                 if (ret)
 144                         goto err;
 145
 146                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 147
 148                 if (!bch2_extents_match(k, old))
 149                         goto nomatch;
 150
 151                 bkey_reassemble(_insert.k, k);
 152                 insert = _insert.k;
 153
 154                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 155                 new = bkey_i_to_extent(_new.k);
 156                 bch2_cut_front(iter.pos, &new->k_i);
 157
 158                 bch2_cut_front(iter.pos,        insert);
 159                 bch2_cut_back(new->k.p,         insert);
 160                 bch2_cut_back(insert->k.p,      &new->k_i);
 161
 162                 /*
 163                  * @old: extent that we read from
 164                  * @insert: key that we're going to update, initialized from
 165                  * extent currently in btree - same as @old unless we raced with
 166                  * other updates
 167                  * @new: extent with new pointers that we'll be adding to @insert
 168                  *
 169                  * Fist, drop rewrite_ptrs from @new:
 170                  */
 171                 i = 0;
 172                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
 173                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 174                             bch2_extent_has_ptr(old, p, bkey_i_to_s_c(insert))) {
 175                                 /*
 176                                  * If we're going to be adding a pointer to the
 177                                  * same device, we have to drop the old one -
 178                                  * otherwise, we can just mark it cached:
 179                                  */
 180                                 if (bch2_bkey_has_device(bkey_i_to_s_c(&new->k_i), p.ptr.dev))
 181                                         bch2_bkey_drop_device_noerror(bkey_i_to_s(insert), p.ptr.dev);
 182                                 else
 183                                         bch2_bkey_mark_dev_cached(bkey_i_to_s(insert), p.ptr.dev);
 184                         }
 185                         i++;
 186                 }
 187
 188
 189                 /* Add new ptrs: */
 190                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 191                         const struct bch_extent_ptr *existing_ptr =
 192                                 bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev);
 193
 194                         if (existing_ptr && existing_ptr->cached) {
 195                                 /*
 196                                  * We're replacing a cached pointer with a non
 197                                  * cached pointer:
 198                                  */
 199                                 bch2_bkey_drop_device_noerror(bkey_i_to_s(insert),
 200                                                               existing_ptr->dev);
 201                         } else if (existing_ptr) {
 202                                 /*
 203                                  * raced with another move op? extent already
 204                                  * has a pointer to the device we just wrote
 205                                  * data to
 206                                  */
 207                                 continue;
 208                         }
 209
 210                         bch2_extent_ptr_decoded_append(insert, &p);
 211                         did_work = true;
 212                 }
 213
 214                 if (!did_work)
 215                         goto nomatch;
 216
 217                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 218                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 219
 220                 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
 221                                                  &should_check_enospc,
 222                                                  &i_sectors_delta,
 223                                                  &disk_sectors_delta);
 224                 if (ret)
 225                         goto err;
 226
 227                 if (disk_sectors_delta > (s64) op->res.sectors) {
 228                         ret = bch2_disk_reservation_add(c, &op->res,
 229                                                 disk_sectors_delta - op->res.sectors,
 230                                                 !should_check_enospc
 231                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 232                         if (ret)
 233                                 goto out;
 234                 }
 235
 236                 next_pos = insert->k.p;
 237
 238                 if (!bkey_eq(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
 239                         ret = insert_snapshot_whiteouts(trans, m->btree_id, k.k->p,
 240                                                         bkey_start_pos(&insert->k));
 241                         if (ret)
 242                                 goto err;
 243                 }
 244
 245                 if (!bkey_eq(insert->k.p, k.k->p)) {
 246                         ret = insert_snapshot_whiteouts(trans, m->btree_id,
 247                                                         k.k->p, insert->k.p);
 248                         if (ret)
 249                                 goto err;
 250                 }
 251
 252                 ret   = bch2_trans_update(trans, &iter, insert,
 253                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 254                         bch2_trans_commit(trans, &op->res,
 255                                 NULL,
 256                                 BTREE_INSERT_NOFAIL|
 257                                 m->data_opts.btree_insert_flags);
 258                 if (!ret) {
 259                         bch2_btree_iter_set_pos(&iter, next_pos);
 260
 261                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 262                         trace_move_extent_finish(&new->k);
 263                 }
 264 err:
 265                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 266                         ret = 0;
 267                 if (ret)
 268                         break;
 269 next:
 270                 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
 271                         bch2_keylist_pop_front(keys);
 272                         if (bch2_keylist_empty(keys))
 273                                 goto out;
 274                 }
 275                 continue;
 276 nomatch:
 277                 if (m->ctxt && m->ctxt->stats) {
 278                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 279                         atomic64_inc(&m->ctxt->stats->keys_raced);
 280                         atomic64_add(k.k->p.offset - iter.pos.offset,
 281                                      &m->ctxt->stats->sectors_raced);
 282                 }
 283
 284                 this_cpu_add(c->counters[BCH_COUNTER_move_extent_fail], new->k.size);
 285                 trace_move_extent_fail(&new->k);
 286
 287                 bch2_btree_iter_advance(&iter);
 288                 goto next;
 289         }
 290 out:
 291         bch2_trans_iter_exit(trans, &iter);
 292         bch2_bkey_buf_exit(&_insert, c);
 293         bch2_bkey_buf_exit(&_new, c);
 294         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 295         return ret;
 296 }
 297
 298 int bch2_data_update_index_update(struct bch_write_op *op)
 299 {
 300         return bch2_trans_run(op->c, __bch2_data_update_index_update(&trans, op));
 301 }
 302
 303 void bch2_data_update_read_done(struct data_update *m,
 304                                 struct bch_extent_crc_unpacked crc)
 305 {
 306         /* write bio must own pages: */
 307         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 308
 309         m->op.crc = crc;
 310         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 311
 312         closure_call(&m->op.cl, bch2_write, NULL, NULL);
 313 }
 314
 315 void bch2_data_update_exit(struct data_update *update)
 316 {
 317         struct bch_fs *c = update->op.c;
 318         struct bkey_ptrs_c ptrs =
 319                 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
 320         const struct bch_extent_ptr *ptr;
 321
 322         bkey_for_each_ptr(ptrs, ptr) {
 323                 if (c->opts.nocow_enabled)
 324                         bch2_bucket_nocow_unlock(&c->nocow_locks,
 325                                                  PTR_BUCKET_POS(c, ptr), 0);
 326                 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 327         }
 328
 329         bch2_bkey_buf_exit(&update->k, c);
 330         bch2_disk_reservation_put(c, &update->op.res);
 331         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 332 }
 333
 334 void bch2_update_unwritten_extent(struct btree_trans *trans,
 335                                   struct data_update *update)
 336 {
 337         struct bch_fs *c = update->op.c;
 338         struct bio *bio = &update->op.wbio.bio;
 339         struct bkey_i_extent *e;
 340         struct write_point *wp;
 341         struct bch_extent_ptr *ptr;
 342         struct closure cl;
 343         struct btree_iter iter;
 344         struct bkey_s_c k;
 345         int ret;
 346
 347         closure_init_stack(&cl);
 348         bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
 349
 350         while (bio_sectors(bio)) {
 351                 unsigned sectors = bio_sectors(bio);
 352
 353                 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
 354                                      BTREE_ITER_SLOTS);
 355                 ret = lockrestart_do(trans, ({
 356                         k = bch2_btree_iter_peek_slot(&iter);
 357                         bkey_err(k);
 358                 }));
 359                 bch2_trans_iter_exit(trans, &iter);
 360
 361                 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
 362                         break;
 363
 364                 e = bkey_extent_init(update->op.insert_keys.top);
 365                 e->k.p = update->op.pos;
 366
 367                 ret = bch2_alloc_sectors_start_trans(trans,
 368                                 update->op.target,
 369                                 false,
 370                                 update->op.write_point,
 371                                 &update->op.devs_have,
 372                                 update->op.nr_replicas,
 373                                 update->op.nr_replicas,
 374                                 update->op.alloc_reserve,
 375                                 0, &cl, &wp);
 376                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
 377                         bch2_trans_unlock(trans);
 378                         closure_sync(&cl);
 379                         continue;
 380                 }
 381
 382                 if (ret)
 383                         return;
 384
 385                 sectors = min(sectors, wp->sectors_free);
 386
 387                 bch2_key_resize(&e->k, sectors);
 388
 389                 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
 390                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 391                 bch2_alloc_sectors_done(c, wp);
 392
 393                 bio_advance(bio, sectors << 9);
 394                 update->op.pos.offset += sectors;
 395
 396                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 397                         ptr->unwritten = true;
 398                 bch2_keylist_push(&update->op.insert_keys);
 399
 400                 ret = __bch2_data_update_index_update(trans, &update->op);
 401
 402                 bch2_open_buckets_put(c, &update->op.open_buckets);
 403
 404                 if (ret)
 405                         break;
 406         }
 407
 408         if ((atomic_read(&cl.remaining) & CLOSURE_REMAINING_MASK) != 1) {
 409                 bch2_trans_unlock(trans);
 410                 closure_sync(&cl);
 411         }
 412 }
 413
 414 int bch2_data_update_init(struct btree_trans *trans,
 415                           struct moving_context *ctxt,
 416                           struct data_update *m,
 417                           struct write_point_specifier wp,
 418                           struct bch_io_opts io_opts,
 419                           struct data_update_opts data_opts,
 420                           enum btree_id btree_id,
 421                           struct bkey_s_c k)
 422 {
 423         struct bch_fs *c = trans->c;
 424         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 425         const union bch_extent_entry *entry;
 426         struct extent_ptr_decoded p;
 427         const struct bch_extent_ptr *ptr;
 428         unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 429         unsigned ptrs_locked = 0;
 430         int ret;
 431
 432         bch2_bkey_buf_init(&m->k);
 433         bch2_bkey_buf_reassemble(&m->k, c, k);
 434         m->btree_id     = btree_id;
 435         m->data_opts    = data_opts;
 436
 437         bch2_write_op_init(&m->op, c, io_opts);
 438         m->op.pos       = bkey_start_pos(k.k);
 439         m->op.version   = k.k->version;
 440         m->op.target    = data_opts.target;
 441         m->op.write_point = wp;
 442         m->op.nr_replicas = 0;
 443         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 444                 BCH_WRITE_PAGES_OWNED|
 445                 BCH_WRITE_DATA_ENCODED|
 446                 BCH_WRITE_MOVE|
 447                 m->data_opts.write_flags;
 448         m->op.compression_type =
 449                 bch2_compression_opt_to_type[io_opts.background_compression ?:
 450                                              io_opts.compression];
 451         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 452                 m->op.alloc_reserve = RESERVE_movinggc;
 453
 454         bkey_for_each_ptr(ptrs, ptr)
 455                 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 456
 457         i = 0;
 458         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 459                 bool locked;
 460
 461                 if (((1U << i) & m->data_opts.rewrite_ptrs)) {
 462                         BUG_ON(p.ptr.cached);
 463
 464                         if (crc_is_compressed(p.crc))
 465                                 reserve_sectors += k.k->size;
 466
 467                         m->op.nr_replicas += bch2_extent_ptr_durability(c, &p);
 468                 } else if (!p.ptr.cached) {
 469                         bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 470                 }
 471
 472                 /*
 473                  * op->csum_type is normally initialized from the fs/file's
 474                  * current options - but if an extent is encrypted, we require
 475                  * that it stays encrypted:
 476                  */
 477                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 478                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 479                         m->op.csum_type = p.crc.csum_type;
 480                 }
 481
 482                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 483                         m->op.incompressible = true;
 484
 485                 if (c->opts.nocow_enabled) {
 486                         if (ctxt) {
 487                                 move_ctxt_wait_event(ctxt, trans,
 488                                                 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
 489                                                                           PTR_BUCKET_POS(c, &p.ptr), 0)) ||
 490                                                 !atomic_read(&ctxt->read_sectors));
 491
 492                                 if (!locked)
 493                                         bch2_bucket_nocow_lock(&c->nocow_locks,
 494                                                                PTR_BUCKET_POS(c, &p.ptr), 0);
 495                         } else {
 496                                 if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
 497                                                                PTR_BUCKET_POS(c, &p.ptr), 0)) {
 498                                         ret = -BCH_ERR_nocow_lock_blocked;
 499                                         goto err;
 500                                 }
 501                         }
 502                         ptrs_locked |= (1U << i);
 503                 }
 504
 505                 i++;
 506         }
 507
 508         if (reserve_sectors) {
 509                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 510                                 m->data_opts.extra_replicas
 511                                 ? 0
 512                                 : BCH_DISK_RESERVATION_NOFAIL);
 513                 if (ret)
 514                         goto err;
 515         }
 516
 517         m->op.nr_replicas += m->data_opts.extra_replicas;
 518         m->op.nr_replicas_required = m->op.nr_replicas;
 519
 520         BUG_ON(!m->op.nr_replicas);
 521
 522         /* Special handling required: */
 523         if (bkey_extent_is_unwritten(k))
 524                 return -BCH_ERR_unwritten_extent_update;
 525         return 0;
 526 err:
 527         i = 0;
 528         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 529                 if ((1U << i) & ptrs_locked)
 530                         bch2_bucket_nocow_unlock(&c->nocow_locks,
 531                                                  PTR_BUCKET_POS(c, &p.ptr), 0);
 532                 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
 533                 i++;
 534         }
 535
 536         bch2_bkey_buf_exit(&m->k, c);
 537         bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
 538         return ret;
 539 }
 540
 541 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 542 {
 543         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 544         const struct bch_extent_ptr *ptr;
 545         unsigned i = 0;
 546
 547         bkey_for_each_ptr(ptrs, ptr) {
 548                 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
 549                         opts->kill_ptrs |= 1U << i;
 550                         opts->rewrite_ptrs ^= 1U << i;
 551                 }
 552
 553                 i++;
 554         }
 555 }