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 "error.h"
  11 #include "extents.h"
  12 #include "io.h"
  13 #include "keylist.h"
  14 #include "move.h"
  15 #include "nocow_locking.h"
  16 #include "subvolume.h"
  17 #include "trace.h"
  18
  19 static void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k)
  20 {
  21         if (trace_move_extent_finish_enabled()) {
  22                 struct printbuf buf = PRINTBUF;
  23
  24                 bch2_bkey_val_to_text(&buf, c, k);
  25                 trace_move_extent_finish(c, buf.buf);
  26                 printbuf_exit(&buf);
  27         }
  28 }
  29
  30 static void trace_move_extent_fail2(struct data_update *m,
  31                          struct bkey_s_c new,
  32                          struct bkey_s_c wrote,
  33                          struct bkey_i *insert,
  34                          const char *msg)
  35 {
  36         struct bch_fs *c = m->op.c;
  37         struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
  38         const union bch_extent_entry *entry;
  39         struct bch_extent_ptr *ptr;
  40         struct extent_ptr_decoded p;
  41         struct printbuf buf = PRINTBUF;
  42         unsigned i, rewrites_found = 0;
  43
  44         if (!trace_move_extent_fail_enabled())
  45                 return;
  46
  47         prt_str(&buf, msg);
  48
  49         if (insert) {
  50                 i = 0;
  51                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
  52                         struct bkey_s new_s;
  53                         new_s.k = (void *) new.k;
  54                         new_s.v = (void *) new.v;
  55
  56                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
  57                             (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
  58                             !ptr->cached)
  59                                 rewrites_found |= 1U << i;
  60                         i++;
  61                 }
  62         }
  63
  64         prt_printf(&buf, "\nrewrite ptrs:   %u%u%u%u",
  65                    (m->data_opts.rewrite_ptrs & (1 << 0)) != 0,
  66                    (m->data_opts.rewrite_ptrs & (1 << 1)) != 0,
  67                    (m->data_opts.rewrite_ptrs & (1 << 2)) != 0,
  68                    (m->data_opts.rewrite_ptrs & (1 << 3)) != 0);
  69
  70         prt_printf(&buf, "\nrewrites found: %u%u%u%u",
  71                    (rewrites_found & (1 << 0)) != 0,
  72                    (rewrites_found & (1 << 1)) != 0,
  73                    (rewrites_found & (1 << 2)) != 0,
  74                    (rewrites_found & (1 << 3)) != 0);
  75
  76         prt_str(&buf, "\nold:    ");
  77         bch2_bkey_val_to_text(&buf, c, old);
  78
  79         prt_str(&buf, "\nnew:    ");
  80         bch2_bkey_val_to_text(&buf, c, new);
  81
  82         prt_str(&buf, "\nwrote:  ");
  83         bch2_bkey_val_to_text(&buf, c, wrote);
  84
  85         if (insert) {
  86                 prt_str(&buf, "\ninsert: ");
  87                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
  88         }
  89
  90         trace_move_extent_fail(c, buf.buf);
  91         printbuf_exit(&buf);
  92 }
  93
  94 static int __bch2_data_update_index_update(struct btree_trans *trans,
  95                                            struct bch_write_op *op)
  96 {
  97         struct bch_fs *c = op->c;
  98         struct btree_iter iter;
  99         struct data_update *m =
 100                 container_of(op, struct data_update, op);
 101         struct keylist *keys = &op->insert_keys;
 102         struct bkey_buf _new, _insert;
 103         int ret = 0;
 104
 105         bch2_bkey_buf_init(&_new);
 106         bch2_bkey_buf_init(&_insert);
 107         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 108
 109         bch2_trans_iter_init(trans, &iter, m->btree_id,
 110                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 111                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 112
 113         while (1) {
 114                 struct bkey_s_c k;
 115                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 116                 struct bkey_i *insert = NULL;
 117                 struct bkey_i_extent *new;
 118                 const union bch_extent_entry *entry_c;
 119                 union bch_extent_entry *entry;
 120                 struct extent_ptr_decoded p;
 121                 struct bch_extent_ptr *ptr;
 122                 const struct bch_extent_ptr *ptr_c;
 123                 struct bpos next_pos;
 124                 bool should_check_enospc;
 125                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 126                 unsigned rewrites_found = 0, durability, i;
 127
 128                 bch2_trans_begin(trans);
 129
 130                 k = bch2_btree_iter_peek_slot(&iter);
 131                 ret = bkey_err(k);
 132                 if (ret)
 133                         goto err;
 134
 135                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 136
 137                 if (!bch2_extents_match(k, old)) {
 138                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i),
 139                                                 NULL, "no match:");
 140                         goto nowork;
 141                 }
 142
 143                 bkey_reassemble(_insert.k, k);
 144                 insert = _insert.k;
 145
 146                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 147                 new = bkey_i_to_extent(_new.k);
 148                 bch2_cut_front(iter.pos, &new->k_i);
 149
 150                 bch2_cut_front(iter.pos,        insert);
 151                 bch2_cut_back(new->k.p,         insert);
 152                 bch2_cut_back(insert->k.p,      &new->k_i);
 153
 154                 /*
 155                  * @old: extent that we read from
 156                  * @insert: key that we're going to update, initialized from
 157                  * extent currently in btree - same as @old unless we raced with
 158                  * other updates
 159                  * @new: extent with new pointers that we'll be adding to @insert
 160                  *
 161                  * Fist, drop rewrite_ptrs from @new:
 162                  */
 163                 i = 0;
 164                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
 165                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 166                             (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
 167                             !ptr->cached) {
 168                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
 169                                 /*
 170                                  * See comment below:
 171                                 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr);
 172                                 */
 173                                 rewrites_found |= 1U << i;
 174                         }
 175                         i++;
 176                 }
 177
 178                 if (m->data_opts.rewrite_ptrs &&
 179                     !rewrites_found &&
 180                     bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) {
 181                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:");
 182                         goto nowork;
 183                 }
 184
 185                 /*
 186                  * A replica that we just wrote might conflict with a replica
 187                  * that we want to keep, due to racing with another move:
 188                  */
 189 restart_drop_conflicting_replicas:
 190                 extent_for_each_ptr(extent_i_to_s(new), ptr)
 191                         if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
 192                             !ptr_c->cached) {
 193                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
 194                                 goto restart_drop_conflicting_replicas;
 195                         }
 196
 197                 if (!bkey_val_u64s(&new->k)) {
 198                         trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:");
 199                         goto nowork;
 200                 }
 201
 202                 /* Now, drop pointers that conflict with what we just wrote: */
 203                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
 204                         if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev)))
 205                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
 206
 207                 durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) +
 208                         bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i));
 209
 210                 /* Now, drop excess replicas: */
 211 restart_drop_extra_replicas:
 212                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
 213                         unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
 214
 215                         if (!p.ptr.cached &&
 216                             durability - ptr_durability >= m->op.opts.data_replicas) {
 217                                 durability -= ptr_durability;
 218                                 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), &entry->ptr);
 219                                 /*
 220                                  * Currently, we're dropping unneeded replicas
 221                                  * instead of marking them as cached, since
 222                                  * cached data in stripe buckets prevents them
 223                                  * from being reused:
 224                                 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr);
 225                                 */
 226                                 goto restart_drop_extra_replicas;
 227                         }
 228                 }
 229
 230                 /* Finally, add the pointers we just wrote: */
 231                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
 232                         bch2_extent_ptr_decoded_append(insert, &p);
 233
 234                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 235                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 236
 237                 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
 238                                                  &should_check_enospc,
 239                                                  &i_sectors_delta,
 240                                                  &disk_sectors_delta);
 241                 if (ret)
 242                         goto err;
 243
 244                 if (disk_sectors_delta > (s64) op->res.sectors) {
 245                         ret = bch2_disk_reservation_add(c, &op->res,
 246                                                 disk_sectors_delta - op->res.sectors,
 247                                                 !should_check_enospc
 248                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 249                         if (ret)
 250                                 goto out;
 251                 }
 252
 253                 next_pos = insert->k.p;
 254
 255                 ret =   bch2_insert_snapshot_whiteouts(trans, m->btree_id,
 256                                                 k.k->p, bkey_start_pos(&insert->k)) ?:
 257                         bch2_insert_snapshot_whiteouts(trans, m->btree_id,
 258                                                 k.k->p, insert->k.p);
 259                 if (ret)
 260                         goto err;
 261
 262                 ret   = bch2_trans_update(trans, &iter, insert,
 263                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 264                         bch2_trans_commit(trans, &op->res,
 265                                 NULL,
 266                                 BTREE_INSERT_NOCHECK_RW|
 267                                 BTREE_INSERT_NOFAIL|
 268                                 m->data_opts.btree_insert_flags);
 269                 if (!ret) {
 270                         bch2_btree_iter_set_pos(&iter, next_pos);
 271
 272                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 273                         trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i));
 274                 }
 275 err:
 276                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 277                         ret = 0;
 278                 if (ret)
 279                         break;
 280 next:
 281                 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
 282                         bch2_keylist_pop_front(keys);
 283                         if (bch2_keylist_empty(keys))
 284                                 goto out;
 285                 }
 286                 continue;
 287 nowork:
 288                 if (m->ctxt && m->ctxt->stats) {
 289                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 290                         atomic64_inc(&m->ctxt->stats->keys_raced);
 291                         atomic64_add(k.k->p.offset - iter.pos.offset,
 292                                      &m->ctxt->stats->sectors_raced);
 293                 }
 294
 295                 this_cpu_inc(c->counters[BCH_COUNTER_move_extent_fail]);
 296
 297                 bch2_btree_iter_advance(&iter);
 298                 goto next;
 299         }
 300 out:
 301         bch2_trans_iter_exit(trans, &iter);
 302         bch2_bkey_buf_exit(&_insert, c);
 303         bch2_bkey_buf_exit(&_new, c);
 304         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 305         return ret;
 306 }
 307
 308 int bch2_data_update_index_update(struct bch_write_op *op)
 309 {
 310         return bch2_trans_run(op->c, __bch2_data_update_index_update(&trans, op));
 311 }
 312
 313 void bch2_data_update_read_done(struct data_update *m,
 314                                 struct bch_extent_crc_unpacked crc)
 315 {
 316         /* write bio must own pages: */
 317         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 318
 319         m->op.crc = crc;
 320         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 321
 322         closure_call(&m->op.cl, bch2_write, NULL, NULL);
 323 }
 324
 325 void bch2_data_update_exit(struct data_update *update)
 326 {
 327         struct bch_fs *c = update->op.c;
 328         struct bkey_ptrs_c ptrs =
 329                 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
 330         const struct bch_extent_ptr *ptr;
 331
 332         bkey_for_each_ptr(ptrs, ptr) {
 333                 if (c->opts.nocow_enabled)
 334                         bch2_bucket_nocow_unlock(&c->nocow_locks,
 335                                                  PTR_BUCKET_POS(c, ptr), 0);
 336                 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 337         }
 338
 339         bch2_bkey_buf_exit(&update->k, c);
 340         bch2_disk_reservation_put(c, &update->op.res);
 341         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 342 }
 343
 344 void bch2_update_unwritten_extent(struct btree_trans *trans,
 345                                   struct data_update *update)
 346 {
 347         struct bch_fs *c = update->op.c;
 348         struct bio *bio = &update->op.wbio.bio;
 349         struct bkey_i_extent *e;
 350         struct write_point *wp;
 351         struct bch_extent_ptr *ptr;
 352         struct closure cl;
 353         struct btree_iter iter;
 354         struct bkey_s_c k;
 355         int ret;
 356
 357         closure_init_stack(&cl);
 358         bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
 359
 360         while (bio_sectors(bio)) {
 361                 unsigned sectors = bio_sectors(bio);
 362
 363                 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
 364                                      BTREE_ITER_SLOTS);
 365                 ret = lockrestart_do(trans, ({
 366                         k = bch2_btree_iter_peek_slot(&iter);
 367                         bkey_err(k);
 368                 }));
 369                 bch2_trans_iter_exit(trans, &iter);
 370
 371                 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
 372                         break;
 373
 374                 e = bkey_extent_init(update->op.insert_keys.top);
 375                 e->k.p = update->op.pos;
 376
 377                 ret = bch2_alloc_sectors_start_trans(trans,
 378                                 update->op.target,
 379                                 false,
 380                                 update->op.write_point,
 381                                 &update->op.devs_have,
 382                                 update->op.nr_replicas,
 383                                 update->op.nr_replicas,
 384                                 update->op.watermark,
 385                                 0, &cl, &wp);
 386                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
 387                         bch2_trans_unlock(trans);
 388                         closure_sync(&cl);
 389                         continue;
 390                 }
 391
 392                 if (ret)
 393                         return;
 394
 395                 sectors = min(sectors, wp->sectors_free);
 396
 397                 bch2_key_resize(&e->k, sectors);
 398
 399                 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
 400                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 401                 bch2_alloc_sectors_done(c, wp);
 402
 403                 bio_advance(bio, sectors << 9);
 404                 update->op.pos.offset += sectors;
 405
 406                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 407                         ptr->unwritten = true;
 408                 bch2_keylist_push(&update->op.insert_keys);
 409
 410                 ret = __bch2_data_update_index_update(trans, &update->op);
 411
 412                 bch2_open_buckets_put(c, &update->op.open_buckets);
 413
 414                 if (ret)
 415                         break;
 416         }
 417
 418         if ((atomic_read(&cl.remaining) & CLOSURE_REMAINING_MASK) != 1) {
 419                 bch2_trans_unlock(trans);
 420                 closure_sync(&cl);
 421         }
 422 }
 423
 424 int bch2_data_update_init(struct btree_trans *trans,
 425                           struct moving_context *ctxt,
 426                           struct data_update *m,
 427                           struct write_point_specifier wp,
 428                           struct bch_io_opts io_opts,
 429                           struct data_update_opts data_opts,
 430                           enum btree_id btree_id,
 431                           struct bkey_s_c k)
 432 {
 433         struct bch_fs *c = trans->c;
 434         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 435         const union bch_extent_entry *entry;
 436         struct extent_ptr_decoded p;
 437         const struct bch_extent_ptr *ptr;
 438         unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 439         unsigned ptrs_locked = 0;
 440         int ret;
 441
 442         bch2_bkey_buf_init(&m->k);
 443         bch2_bkey_buf_reassemble(&m->k, c, k);
 444         m->btree_id     = btree_id;
 445         m->data_opts    = data_opts;
 446
 447         bch2_write_op_init(&m->op, c, io_opts);
 448         m->op.pos       = bkey_start_pos(k.k);
 449         m->op.version   = k.k->version;
 450         m->op.target    = data_opts.target;
 451         m->op.write_point = wp;
 452         m->op.nr_replicas = 0;
 453         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 454                 BCH_WRITE_PAGES_OWNED|
 455                 BCH_WRITE_DATA_ENCODED|
 456                 BCH_WRITE_MOVE|
 457                 m->data_opts.write_flags;
 458         m->op.compression_opt   = io_opts.background_compression ?: io_opts.compression;
 459         m->op.watermark         = m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
 460
 461         bkey_for_each_ptr(ptrs, ptr)
 462                 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 463
 464         i = 0;
 465         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 466                 bool locked;
 467
 468                 if (((1U << i) & m->data_opts.rewrite_ptrs)) {
 469                         BUG_ON(p.ptr.cached);
 470
 471                         if (crc_is_compressed(p.crc))
 472                                 reserve_sectors += k.k->size;
 473
 474                         m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
 475                 } else if (!p.ptr.cached) {
 476                         bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 477                 }
 478
 479                 /*
 480                  * op->csum_type is normally initialized from the fs/file's
 481                  * current options - but if an extent is encrypted, we require
 482                  * that it stays encrypted:
 483                  */
 484                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 485                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 486                         m->op.csum_type = p.crc.csum_type;
 487                 }
 488
 489                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 490                         m->op.incompressible = true;
 491
 492                 if (c->opts.nocow_enabled) {
 493                         if (ctxt) {
 494                                 move_ctxt_wait_event(ctxt, trans,
 495                                                 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
 496                                                                           PTR_BUCKET_POS(c, &p.ptr), 0)) ||
 497                                                 !atomic_read(&ctxt->read_sectors));
 498
 499                                 if (!locked)
 500                                         bch2_bucket_nocow_lock(&c->nocow_locks,
 501                                                                PTR_BUCKET_POS(c, &p.ptr), 0);
 502                         } else {
 503                                 if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
 504                                                                PTR_BUCKET_POS(c, &p.ptr), 0)) {
 505                                         ret = -BCH_ERR_nocow_lock_blocked;
 506                                         goto err;
 507                                 }
 508                         }
 509                         ptrs_locked |= (1U << i);
 510                 }
 511
 512                 i++;
 513         }
 514
 515         if (reserve_sectors) {
 516                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 517                                 m->data_opts.extra_replicas
 518                                 ? 0
 519                                 : BCH_DISK_RESERVATION_NOFAIL);
 520                 if (ret)
 521                         goto err;
 522         }
 523
 524         m->op.nr_replicas += m->data_opts.extra_replicas;
 525         m->op.nr_replicas_required = m->op.nr_replicas;
 526
 527         BUG_ON(!m->op.nr_replicas);
 528
 529         /* Special handling required: */
 530         if (bkey_extent_is_unwritten(k))
 531                 return -BCH_ERR_unwritten_extent_update;
 532         return 0;
 533 err:
 534         i = 0;
 535         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 536                 if ((1U << i) & ptrs_locked)
 537                         bch2_bucket_nocow_unlock(&c->nocow_locks,
 538                                                  PTR_BUCKET_POS(c, &p.ptr), 0);
 539                 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
 540                 i++;
 541         }
 542
 543         bch2_bkey_buf_exit(&m->k, c);
 544         bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
 545         return ret;
 546 }
 547
 548 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 549 {
 550         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 551         const struct bch_extent_ptr *ptr;
 552         unsigned i = 0;
 553
 554         bkey_for_each_ptr(ptrs, ptr) {
 555                 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
 556                         opts->kill_ptrs |= 1U << i;
 557                         opts->rewrite_ptrs ^= 1U << i;
 558                 }
 559
 560                 i++;
 561         }
 562 }