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, update_iter;
  26         struct bkey_s_c k;
  27         snapshot_id_list s;
  28         int ret;
  29
  30         if (!btree_type_has_snapshots(id))
  31                 return 0;
  32
  33         darray_init(&s);
  34
  35         if (bkey_eq(old_pos, new_pos))
  36                 return 0;
  37
  38         if (!snapshot_t(c, old_pos.snapshot)->children[0])
  39                 return 0;
  40
  41         bch2_trans_iter_init(trans, &iter, id, old_pos,
  42                              BTREE_ITER_NOT_EXTENTS|
  43                              BTREE_ITER_ALL_SNAPSHOTS);
  44         while (1) {
  45                 k = bch2_btree_iter_prev(&iter);
  46                 ret = bkey_err(k);
  47                 if (ret)
  48                         break;
  49
  50                 if (!bkey_eq(old_pos, k.k->p))
  51                         break;
  52
  53                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot)) {
  54                         struct bkey_i *update;
  55
  56                         if (snapshot_list_has_ancestor(c, &s, k.k->p.snapshot))
  57                                 continue;
  58
  59                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
  60
  61                         ret = PTR_ERR_OR_ZERO(update);
  62                         if (ret)
  63                                 break;
  64
  65                         bkey_init(&update->k);
  66                         update->k.p = new_pos;
  67                         update->k.p.snapshot = k.k->p.snapshot;
  68
  69                         bch2_trans_iter_init(trans, &update_iter, id, update->k.p,
  70                                              BTREE_ITER_NOT_EXTENTS|
  71                                              BTREE_ITER_ALL_SNAPSHOTS|
  72                                              BTREE_ITER_INTENT);
  73                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
  74                                 bch2_trans_update(trans, &update_iter, update,
  75                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
  76                         bch2_trans_iter_exit(trans, &update_iter);
  77                         if (ret)
  78                                 break;
  79
  80                         ret = snapshot_list_add(c, &s, k.k->p.snapshot);
  81                         if (ret)
  82                                 break;
  83                 }
  84         }
  85         bch2_trans_iter_exit(trans, &iter);
  86         darray_exit(&s);
  87
  88         return ret;
  89 }
  90
  91 static void bch2_bkey_mark_dev_cached(struct bkey_s k, unsigned dev)
  92 {
  93         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
  94         struct bch_extent_ptr *ptr;
  95
  96         bkey_for_each_ptr(ptrs, ptr)
  97                 if (ptr->dev == dev)
  98                         ptr->cached = true;
  99 }
 100
 101 static int __bch2_data_update_index_update(struct btree_trans *trans,
 102                                            struct bch_write_op *op)
 103 {
 104         struct bch_fs *c = op->c;
 105         struct btree_iter iter;
 106         struct data_update *m =
 107                 container_of(op, struct data_update, op);
 108         struct keylist *keys = &op->insert_keys;
 109         struct bkey_buf _new, _insert;
 110         int ret = 0;
 111
 112         bch2_bkey_buf_init(&_new);
 113         bch2_bkey_buf_init(&_insert);
 114         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 115
 116         bch2_trans_iter_init(trans, &iter, m->btree_id,
 117                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 118                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 119
 120         while (1) {
 121                 struct bkey_s_c k;
 122                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 123                 struct bkey_i *insert;
 124                 struct bkey_i_extent *new;
 125                 const union bch_extent_entry *entry;
 126                 struct extent_ptr_decoded p;
 127                 struct bpos next_pos;
 128                 bool did_work = false;
 129                 bool should_check_enospc;
 130                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 131                 unsigned i;
 132
 133                 bch2_trans_begin(trans);
 134
 135                 k = bch2_btree_iter_peek_slot(&iter);
 136                 ret = bkey_err(k);
 137                 if (ret)
 138                         goto err;
 139
 140                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 141
 142                 if (!bch2_extents_match(k, old))
 143                         goto nomatch;
 144
 145                 bkey_reassemble(_insert.k, k);
 146                 insert = _insert.k;
 147
 148                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 149                 new = bkey_i_to_extent(_new.k);
 150                 bch2_cut_front(iter.pos, &new->k_i);
 151
 152                 bch2_cut_front(iter.pos,        insert);
 153                 bch2_cut_back(new->k.p,         insert);
 154                 bch2_cut_back(insert->k.p,      &new->k_i);
 155
 156                 /*
 157                  * @old: extent that we read from
 158                  * @insert: key that we're going to update, initialized from
 159                  * extent currently in btree - same as @old unless we raced with
 160                  * other updates
 161                  * @new: extent with new pointers that we'll be adding to @insert
 162                  *
 163                  * Fist, drop rewrite_ptrs from @new:
 164                  */
 165                 i = 0;
 166                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
 167                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 168                             bch2_extent_has_ptr(old, p, bkey_i_to_s_c(insert))) {
 169                                 /*
 170                                  * If we're going to be adding a pointer to the
 171                                  * same device, we have to drop the old one -
 172                                  * otherwise, we can just mark it cached:
 173                                  */
 174                                 if (bch2_bkey_has_device(bkey_i_to_s_c(&new->k_i), p.ptr.dev))
 175                                         bch2_bkey_drop_device_noerror(bkey_i_to_s(insert), p.ptr.dev);
 176                                 else
 177                                         bch2_bkey_mark_dev_cached(bkey_i_to_s(insert), p.ptr.dev);
 178                         }
 179                         i++;
 180                 }
 181
 182
 183                 /* Add new ptrs: */
 184                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 185                         const struct bch_extent_ptr *existing_ptr =
 186                                 bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev);
 187
 188                         if (existing_ptr && existing_ptr->cached) {
 189                                 /*
 190                                  * We're replacing a cached pointer with a non
 191                                  * cached pointer:
 192                                  */
 193                                 bch2_bkey_drop_device_noerror(bkey_i_to_s(insert),
 194                                                               existing_ptr->dev);
 195                         } else if (existing_ptr) {
 196                                 /*
 197                                  * raced with another move op? extent already
 198                                  * has a pointer to the device we just wrote
 199                                  * data to
 200                                  */
 201                                 continue;
 202                         }
 203
 204                         bch2_extent_ptr_decoded_append(insert, &p);
 205                         did_work = true;
 206                 }
 207
 208                 if (!did_work)
 209                         goto nomatch;
 210
 211                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 212                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 213
 214                 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
 215                                                  &should_check_enospc,
 216                                                  &i_sectors_delta,
 217                                                  &disk_sectors_delta);
 218                 if (ret)
 219                         goto err;
 220
 221                 if (disk_sectors_delta > (s64) op->res.sectors) {
 222                         ret = bch2_disk_reservation_add(c, &op->res,
 223                                                 disk_sectors_delta - op->res.sectors,
 224                                                 !should_check_enospc
 225                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 226                         if (ret)
 227                                 goto out;
 228                 }
 229
 230                 next_pos = insert->k.p;
 231
 232                 ret   = insert_snapshot_whiteouts(trans, m->btree_id,
 233                                                   k.k->p, insert->k.p) ?:
 234                         bch2_trans_update(trans, &iter, insert,
 235                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 236                         bch2_trans_commit(trans, &op->res,
 237                                 NULL,
 238                                 BTREE_INSERT_NOFAIL|
 239                                 m->data_opts.btree_insert_flags);
 240                 if (!ret) {
 241                         bch2_btree_iter_set_pos(&iter, next_pos);
 242
 243                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 244                         trace_move_extent_finish(&new->k);
 245                 }
 246 err:
 247                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 248                         ret = 0;
 249                 if (ret)
 250                         break;
 251 next:
 252                 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
 253                         bch2_keylist_pop_front(keys);
 254                         if (bch2_keylist_empty(keys))
 255                                 goto out;
 256                 }
 257                 continue;
 258 nomatch:
 259                 if (m->ctxt) {
 260                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 261                         atomic64_inc(&m->ctxt->stats->keys_raced);
 262                         atomic64_add(k.k->p.offset - iter.pos.offset,
 263                                      &m->ctxt->stats->sectors_raced);
 264                 }
 265
 266                 this_cpu_add(c->counters[BCH_COUNTER_move_extent_fail], new->k.size);
 267                 trace_move_extent_fail(&new->k);
 268
 269                 bch2_btree_iter_advance(&iter);
 270                 goto next;
 271         }
 272 out:
 273         bch2_trans_iter_exit(trans, &iter);
 274         bch2_bkey_buf_exit(&_insert, c);
 275         bch2_bkey_buf_exit(&_new, c);
 276         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 277         return ret;
 278 }
 279
 280 int bch2_data_update_index_update(struct bch_write_op *op)
 281 {
 282         struct bch_fs *c = op->c;
 283         struct btree_trans trans;
 284         int ret;
 285
 286         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 287         ret = __bch2_data_update_index_update(&trans, op);
 288         bch2_trans_exit(&trans);
 289
 290         return ret;
 291 }
 292
 293 void bch2_data_update_read_done(struct data_update *m,
 294                                 struct bch_extent_crc_unpacked crc)
 295 {
 296         /* write bio must own pages: */
 297         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 298
 299         m->op.crc = crc;
 300         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 301
 302         closure_call(&m->op.cl, bch2_write, NULL, NULL);
 303 }
 304
 305 void bch2_data_update_exit(struct data_update *update)
 306 {
 307         struct bch_fs *c = update->op.c;
 308         struct bkey_ptrs_c ptrs =
 309                 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
 310         const struct bch_extent_ptr *ptr;
 311
 312         bkey_for_each_ptr(ptrs, ptr) {
 313                 bch2_bucket_nocow_unlock(&c->nocow_locks,
 314                                          PTR_BUCKET_POS(c, ptr), 0);
 315                 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 316         }
 317
 318         bch2_bkey_buf_exit(&update->k, c);
 319         bch2_disk_reservation_put(c, &update->op.res);
 320         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 321 }
 322
 323 void bch2_update_unwritten_extent(struct btree_trans *trans,
 324                                   struct data_update *update)
 325 {
 326         struct bch_fs *c = update->op.c;
 327         struct bio *bio = &update->op.wbio.bio;
 328         struct bkey_i_extent *e;
 329         struct write_point *wp;
 330         struct bch_extent_ptr *ptr;
 331         struct closure cl;
 332         struct btree_iter iter;
 333         struct bkey_s_c k;
 334         int ret;
 335
 336         closure_init_stack(&cl);
 337         bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
 338
 339         while (bio_sectors(bio)) {
 340                 unsigned sectors = bio_sectors(bio);
 341
 342                 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
 343                                      BTREE_ITER_SLOTS);
 344                 ret = lockrestart_do(trans, ({
 345                         k = bch2_btree_iter_peek_slot(&iter);
 346                         bkey_err(k);
 347                 }));
 348                 bch2_trans_iter_exit(trans, &iter);
 349
 350                 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
 351                         break;
 352
 353                 e = bkey_extent_init(update->op.insert_keys.top);
 354                 e->k.p = update->op.pos;
 355
 356                 ret = bch2_alloc_sectors_start_trans(trans,
 357                                 update->op.target,
 358                                 false,
 359                                 update->op.write_point,
 360                                 &update->op.devs_have,
 361                                 update->op.nr_replicas,
 362                                 update->op.nr_replicas,
 363                                 update->op.alloc_reserve,
 364                                 0, &cl, &wp);
 365                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
 366                         bch2_trans_unlock(trans);
 367                         closure_sync(&cl);
 368                         continue;
 369                 }
 370
 371                 if (ret)
 372                         return;
 373
 374                 sectors = min(sectors, wp->sectors_free);
 375
 376                 bch2_key_resize(&e->k, sectors);
 377
 378                 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
 379                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 380                 bch2_alloc_sectors_done(c, wp);
 381
 382                 bio_advance(bio, sectors << 9);
 383                 update->op.pos.offset += sectors;
 384
 385                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 386                         ptr->unwritten = true;
 387                 bch2_keylist_push(&update->op.insert_keys);
 388
 389                 ret = __bch2_data_update_index_update(trans, &update->op);
 390
 391                 bch2_open_buckets_put(c, &update->op.open_buckets);
 392
 393                 if (ret)
 394                         break;
 395         }
 396
 397         if ((atomic_read(&cl.remaining) & CLOSURE_REMAINING_MASK) != 1) {
 398                 bch2_trans_unlock(trans);
 399                 closure_sync(&cl);
 400         }
 401 }
 402
 403 int bch2_data_update_init(struct btree_trans *trans,
 404                           struct moving_context *ctxt,
 405                           struct data_update *m,
 406                           struct write_point_specifier wp,
 407                           struct bch_io_opts io_opts,
 408                           struct data_update_opts data_opts,
 409                           enum btree_id btree_id,
 410                           struct bkey_s_c k)
 411 {
 412         struct bch_fs *c = trans->c;
 413         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 414         const union bch_extent_entry *entry;
 415         struct extent_ptr_decoded p;
 416         const struct bch_extent_ptr *ptr;
 417         unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 418         unsigned int ptrs_locked = 0;
 419         int ret;
 420
 421         bch2_bkey_buf_init(&m->k);
 422         bch2_bkey_buf_reassemble(&m->k, c, k);
 423         m->btree_id     = btree_id;
 424         m->data_opts    = data_opts;
 425
 426         bch2_write_op_init(&m->op, c, io_opts);
 427         m->op.pos       = bkey_start_pos(k.k);
 428         m->op.version   = k.k->version;
 429         m->op.target    = data_opts.target;
 430         m->op.write_point = wp;
 431         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 432                 BCH_WRITE_PAGES_OWNED|
 433                 BCH_WRITE_DATA_ENCODED|
 434                 BCH_WRITE_MOVE|
 435                 m->data_opts.write_flags;
 436         m->op.compression_type =
 437                 bch2_compression_opt_to_type[io_opts.background_compression ?:
 438                                              io_opts.compression];
 439         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 440                 m->op.alloc_reserve = RESERVE_movinggc;
 441
 442         bkey_for_each_ptr(ptrs, ptr)
 443                 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 444
 445         i = 0;
 446         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 447                 bool locked;
 448
 449                 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 450                     p.ptr.cached)
 451                         BUG();
 452
 453                 if (!((1U << i) & m->data_opts.rewrite_ptrs) &&
 454                     !p.ptr.cached)
 455                         bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 456
 457                 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 458                     crc_is_compressed(p.crc))
 459                         reserve_sectors += k.k->size;
 460
 461                 /*
 462                  * op->csum_type is normally initialized from the fs/file's
 463                  * current options - but if an extent is encrypted, we require
 464                  * that it stays encrypted:
 465                  */
 466                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 467                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 468                         m->op.csum_type = p.crc.csum_type;
 469                 }
 470
 471                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 472                         m->op.incompressible = true;
 473
 474                 if (ctxt) {
 475                         move_ctxt_wait_event(ctxt, trans,
 476                                         (locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
 477                                                                   PTR_BUCKET_POS(c, &p.ptr), 0)) ||
 478                                         !atomic_read(&ctxt->read_sectors));
 479
 480                         if (!locked)
 481                                 bch2_bucket_nocow_lock(&c->nocow_locks,
 482                                                        PTR_BUCKET_POS(c, &p.ptr), 0);
 483                 } else {
 484                         if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
 485                                                        PTR_BUCKET_POS(c, &p.ptr), 0)) {
 486                                 ret = -BCH_ERR_nocow_lock_blocked;
 487                                 goto err;
 488                         }
 489                 }
 490                 ptrs_locked |= (1U << i);
 491                 i++;
 492         }
 493
 494         if (reserve_sectors) {
 495                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 496                                 m->data_opts.extra_replicas
 497                                 ? 0
 498                                 : BCH_DISK_RESERVATION_NOFAIL);
 499                 if (ret)
 500                         goto err;
 501         }
 502
 503         m->op.nr_replicas = m->op.nr_replicas_required =
 504                 hweight32(m->data_opts.rewrite_ptrs) + m->data_opts.extra_replicas;
 505
 506         BUG_ON(!m->op.nr_replicas);
 507
 508         /* Special handling required: */
 509         if (bkey_extent_is_unwritten(k))
 510                 return -BCH_ERR_unwritten_extent_update;
 511         return 0;
 512 err:
 513         i = 0;
 514         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 515                 if ((1U << i) & ptrs_locked)
 516                         bch2_bucket_nocow_unlock(&c->nocow_locks,
 517                                                  PTR_BUCKET_POS(c, &p.ptr), 0);
 518                 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
 519                 i++;
 520         }
 521
 522         bch2_bkey_buf_exit(&m->k, c);
 523         bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
 524         return ret;
 525 }
 526
 527 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 528 {
 529         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 530         const struct bch_extent_ptr *ptr;
 531         unsigned i = 0;
 532
 533         bkey_for_each_ptr(ptrs, ptr) {
 534                 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
 535                         opts->kill_ptrs |= 1U << i;
 536                         opts->rewrite_ptrs ^= 1U << i;
 537                 }
 538
 539                 i++;
 540         }
 541 }