fs/bcachefs/btree_update_interior.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 "bkey_methods.h"
   7 #include "btree_cache.h"
   8 #include "btree_gc.h"
   9 #include "btree_journal_iter.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "btree_io.h"
  13 #include "btree_iter.h"
  14 #include "btree_locking.h"
  15 #include "buckets.h"
  16 #include "clock.h"
  17 #include "error.h"
  18 #include "extents.h"
  19 #include "journal.h"
  20 #include "journal_reclaim.h"
  21 #include "keylist.h"
  22 #include "recovery_passes.h"
  23 #include "replicas.h"
  24 #include "sb-members.h"
  25 #include "super-io.h"
  26 #include "trace.h"
  27
  28 #include <linux/random.h>
  29
  30 static const char * const bch2_btree_update_modes[] = {
  31 #define x(t) #t,
  32         BTREE_UPDATE_MODES()
  33 #undef x
  34         NULL
  35 };
  36
  37 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
  38                                   btree_path_idx_t, struct btree *, struct keylist *);
  39 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  40
  41 /*
  42  * Verify that child nodes correctly span parent node's range:
  43  */
  44 int bch2_btree_node_check_topology(struct btree_trans *trans, struct btree *b)
  45 {
  46         struct bch_fs *c = trans->c;
  47         struct bpos node_min = b->key.k.type == KEY_TYPE_btree_ptr_v2
  48                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.min_key
  49                 : b->data->min_key;
  50         struct btree_and_journal_iter iter;
  51         struct bkey_s_c k;
  52         struct printbuf buf = PRINTBUF;
  53         struct bkey_buf prev;
  54         int ret = 0;
  55
  56         BUG_ON(b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
  57                !bpos_eq(bkey_i_to_btree_ptr_v2(&b->key)->v.min_key,
  58                         b->data->min_key));
  59
  60         if (b == btree_node_root(c, b)) {
  61                 if (!bpos_eq(b->data->min_key, POS_MIN)) {
  62                         printbuf_reset(&buf);
  63                         bch2_bpos_to_text(&buf, b->data->min_key);
  64                         need_fsck_err(c, btree_root_bad_min_key,
  65                                       "btree root with incorrect min_key: %s", buf.buf);
  66                         goto topology_repair;
  67                 }
  68
  69                 if (!bpos_eq(b->data->max_key, SPOS_MAX)) {
  70                         printbuf_reset(&buf);
  71                         bch2_bpos_to_text(&buf, b->data->max_key);
  72                         need_fsck_err(c, btree_root_bad_max_key,
  73                                       "btree root with incorrect max_key: %s", buf.buf);
  74                         goto topology_repair;
  75                 }
  76         }
  77
  78         if (!b->c.level)
  79                 return 0;
  80
  81         bch2_bkey_buf_init(&prev);
  82         bkey_init(&prev.k->k);
  83         bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
  84
  85         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
  86                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  87                         goto out;
  88
  89                 struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
  90
  91                 struct bpos expected_min = bkey_deleted(&prev.k->k)
  92                         ? node_min
  93                         : bpos_successor(prev.k->k.p);
  94
  95                 if (!bpos_eq(expected_min, bp.v->min_key)) {
  96                         bch2_topology_error(c);
  97
  98                         printbuf_reset(&buf);
  99                         prt_str(&buf, "end of prev node doesn't match start of next node\n"),
 100                         prt_printf(&buf, "  in btree %s level %u node ",
 101                                    bch2_btree_id_str(b->c.btree_id), b->c.level);
 102                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 103                         prt_str(&buf, "\n  prev ");
 104                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(prev.k));
 105                         prt_str(&buf, "\n  next ");
 106                         bch2_bkey_val_to_text(&buf, c, k);
 107
 108                         need_fsck_err(c, btree_node_topology_bad_min_key, "%s", buf.buf);
 109                         goto topology_repair;
 110                 }
 111
 112                 bch2_bkey_buf_reassemble(&prev, c, k);
 113                 bch2_btree_and_journal_iter_advance(&iter);
 114         }
 115
 116         if (bkey_deleted(&prev.k->k)) {
 117                 bch2_topology_error(c);
 118
 119                 printbuf_reset(&buf);
 120                 prt_str(&buf, "empty interior node\n");
 121                 prt_printf(&buf, "  in btree %s level %u node ",
 122                            bch2_btree_id_str(b->c.btree_id), b->c.level);
 123                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 124
 125                 need_fsck_err(c, btree_node_topology_empty_interior_node, "%s", buf.buf);
 126                 goto topology_repair;
 127         } else if (!bpos_eq(prev.k->k.p, b->key.k.p)) {
 128                 bch2_topology_error(c);
 129
 130                 printbuf_reset(&buf);
 131                 prt_str(&buf, "last child node doesn't end at end of parent node\n");
 132                 prt_printf(&buf, "  in btree %s level %u node ",
 133                            bch2_btree_id_str(b->c.btree_id), b->c.level);
 134                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 135                 prt_str(&buf, "\n  last key ");
 136                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(prev.k));
 137
 138                 need_fsck_err(c, btree_node_topology_bad_max_key, "%s", buf.buf);
 139                 goto topology_repair;
 140         }
 141 out:
 142 fsck_err:
 143         bch2_btree_and_journal_iter_exit(&iter);
 144         bch2_bkey_buf_exit(&prev, c);
 145         printbuf_exit(&buf);
 146         return ret;
 147 topology_repair:
 148         if ((c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology)) &&
 149             c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology) {
 150                 bch2_inconsistent_error(c);
 151                 ret = -BCH_ERR_btree_need_topology_repair;
 152         } else {
 153                 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 154         }
 155         goto out;
 156 }
 157
 158 /* Calculate ideal packed bkey format for new btree nodes: */
 159
 160 static void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
 161 {
 162         struct bkey_packed *k;
 163         struct bkey uk;
 164
 165         for_each_bset(b, t)
 166                 bset_tree_for_each_key(b, t, k)
 167                         if (!bkey_deleted(k)) {
 168                                 uk = bkey_unpack_key(b, k);
 169                                 bch2_bkey_format_add_key(s, &uk);
 170                         }
 171 }
 172
 173 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 174 {
 175         struct bkey_format_state s;
 176
 177         bch2_bkey_format_init(&s);
 178         bch2_bkey_format_add_pos(&s, b->data->min_key);
 179         bch2_bkey_format_add_pos(&s, b->data->max_key);
 180         __bch2_btree_calc_format(&s, b);
 181
 182         return bch2_bkey_format_done(&s);
 183 }
 184
 185 static size_t btree_node_u64s_with_format(struct btree_nr_keys nr,
 186                                           struct bkey_format *old_f,
 187                                           struct bkey_format *new_f)
 188 {
 189         /* stupid integer promotion rules */
 190         ssize_t delta =
 191             (((int) new_f->key_u64s - old_f->key_u64s) *
 192              (int) nr.packed_keys) +
 193             (((int) new_f->key_u64s - BKEY_U64s) *
 194              (int) nr.unpacked_keys);
 195
 196         BUG_ON(delta + nr.live_u64s < 0);
 197
 198         return nr.live_u64s + delta;
 199 }
 200
 201 /**
 202  * bch2_btree_node_format_fits - check if we could rewrite node with a new format
 203  *
 204  * @c:          filesystem handle
 205  * @b:          btree node to rewrite
 206  * @nr:         number of keys for new node (i.e. b->nr)
 207  * @new_f:      bkey format to translate keys to
 208  *
 209  * Returns: true if all re-packed keys will be able to fit in a new node.
 210  *
 211  * Assumes all keys will successfully pack with the new format.
 212  */
 213 static bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 214                                  struct btree_nr_keys nr,
 215                                  struct bkey_format *new_f)
 216 {
 217         size_t u64s = btree_node_u64s_with_format(nr, &b->format, new_f);
 218
 219         return __vstruct_bytes(struct btree_node, u64s) < btree_buf_bytes(b);
 220 }
 221
 222 /* Btree node freeing/allocation: */
 223
 224 static void __btree_node_free(struct btree_trans *trans, struct btree *b)
 225 {
 226         struct bch_fs *c = trans->c;
 227
 228         trace_and_count(c, btree_node_free, trans, b);
 229
 230         BUG_ON(btree_node_write_blocked(b));
 231         BUG_ON(btree_node_dirty(b));
 232         BUG_ON(btree_node_need_write(b));
 233         BUG_ON(b == btree_node_root(c, b));
 234         BUG_ON(b->ob.nr);
 235         BUG_ON(!list_empty(&b->write_blocked));
 236         BUG_ON(b->will_make_reachable);
 237
 238         clear_btree_node_noevict(b);
 239
 240         mutex_lock(&c->btree_cache.lock);
 241         list_move(&b->list, &c->btree_cache.freeable);
 242         mutex_unlock(&c->btree_cache.lock);
 243 }
 244
 245 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 246                                        struct btree_path *path,
 247                                        struct btree *b)
 248 {
 249         struct bch_fs *c = trans->c;
 250         unsigned i, level = b->c.level;
 251
 252         bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 253         bch2_btree_node_hash_remove(&c->btree_cache, b);
 254         __btree_node_free(trans, b);
 255         six_unlock_write(&b->c.lock);
 256         mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 257
 258         trans_for_each_path(trans, path, i)
 259                 if (path->l[level].b == b) {
 260                         btree_node_unlock(trans, path, level);
 261                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 262                 }
 263 }
 264
 265 static void bch2_btree_node_free_never_used(struct btree_update *as,
 266                                             struct btree_trans *trans,
 267                                             struct btree *b)
 268 {
 269         struct bch_fs *c = as->c;
 270         struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
 271         struct btree_path *path;
 272         unsigned i, level = b->c.level;
 273
 274         BUG_ON(!list_empty(&b->write_blocked));
 275         BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
 276
 277         b->will_make_reachable = 0;
 278         closure_put(&as->cl);
 279
 280         clear_btree_node_will_make_reachable(b);
 281         clear_btree_node_accessed(b);
 282         clear_btree_node_dirty_acct(c, b);
 283         clear_btree_node_need_write(b);
 284
 285         mutex_lock(&c->btree_cache.lock);
 286         list_del_init(&b->list);
 287         bch2_btree_node_hash_remove(&c->btree_cache, b);
 288         mutex_unlock(&c->btree_cache.lock);
 289
 290         BUG_ON(p->nr >= ARRAY_SIZE(p->b));
 291         p->b[p->nr++] = b;
 292
 293         six_unlock_intent(&b->c.lock);
 294
 295         trans_for_each_path(trans, path, i)
 296                 if (path->l[level].b == b) {
 297                         btree_node_unlock(trans, path, level);
 298                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 299                 }
 300 }
 301
 302 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 303                                              struct disk_reservation *res,
 304                                              struct closure *cl,
 305                                              bool interior_node,
 306                                              unsigned flags)
 307 {
 308         struct bch_fs *c = trans->c;
 309         struct write_point *wp;
 310         struct btree *b;
 311         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 312         struct open_buckets obs = { .nr = 0 };
 313         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 314         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 315         unsigned nr_reserve = watermark < BCH_WATERMARK_reclaim
 316                 ? BTREE_NODE_RESERVE
 317                 : 0;
 318         int ret;
 319
 320         mutex_lock(&c->btree_reserve_cache_lock);
 321         if (c->btree_reserve_cache_nr > nr_reserve) {
 322                 struct btree_alloc *a =
 323                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 324
 325                 obs = a->ob;
 326                 bkey_copy(&tmp.k, &a->k);
 327                 mutex_unlock(&c->btree_reserve_cache_lock);
 328                 goto mem_alloc;
 329         }
 330         mutex_unlock(&c->btree_reserve_cache_lock);
 331
 332 retry:
 333         ret = bch2_alloc_sectors_start_trans(trans,
 334                                       c->opts.metadata_target ?:
 335                                       c->opts.foreground_target,
 336                                       0,
 337                                       writepoint_ptr(&c->btree_write_point),
 338                                       &devs_have,
 339                                       res->nr_replicas,
 340                                       min(res->nr_replicas,
 341                                           c->opts.metadata_replicas_required),
 342                                       watermark, 0, cl, &wp);
 343         if (unlikely(ret))
 344                 return ERR_PTR(ret);
 345
 346         if (wp->sectors_free < btree_sectors(c)) {
 347                 struct open_bucket *ob;
 348                 unsigned i;
 349
 350                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 351                         if (ob->sectors_free < btree_sectors(c))
 352                                 ob->sectors_free = 0;
 353
 354                 bch2_alloc_sectors_done(c, wp);
 355                 goto retry;
 356         }
 357
 358         bkey_btree_ptr_v2_init(&tmp.k);
 359         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 360
 361         bch2_open_bucket_get(c, wp, &obs);
 362         bch2_alloc_sectors_done(c, wp);
 363 mem_alloc:
 364         b = bch2_btree_node_mem_alloc(trans, interior_node);
 365         six_unlock_write(&b->c.lock);
 366         six_unlock_intent(&b->c.lock);
 367
 368         /* we hold cannibalize_lock: */
 369         BUG_ON(IS_ERR(b));
 370         BUG_ON(b->ob.nr);
 371
 372         bkey_copy(&b->key, &tmp.k);
 373         b->ob = obs;
 374
 375         return b;
 376 }
 377
 378 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 379                                            struct btree_trans *trans,
 380                                            unsigned level)
 381 {
 382         struct bch_fs *c = as->c;
 383         struct btree *b;
 384         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 385         int ret;
 386
 387         BUG_ON(level >= BTREE_MAX_DEPTH);
 388         BUG_ON(!p->nr);
 389
 390         b = p->b[--p->nr];
 391
 392         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 393         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 394
 395         set_btree_node_accessed(b);
 396         set_btree_node_dirty_acct(c, b);
 397         set_btree_node_need_write(b);
 398
 399         bch2_bset_init_first(b, &b->data->keys);
 400         b->c.level      = level;
 401         b->c.btree_id   = as->btree_id;
 402         b->version_ondisk = c->sb.version;
 403
 404         memset(&b->nr, 0, sizeof(b->nr));
 405         b->data->magic = cpu_to_le64(bset_magic(c));
 406         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 407         b->data->flags = 0;
 408         SET_BTREE_NODE_ID(b->data, as->btree_id);
 409         SET_BTREE_NODE_LEVEL(b->data, level);
 410
 411         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 412                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 413
 414                 bp->v.mem_ptr           = 0;
 415                 bp->v.seq               = b->data->keys.seq;
 416                 bp->v.sectors_written   = 0;
 417         }
 418
 419         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 420
 421         bch2_btree_build_aux_trees(b);
 422
 423         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 424         BUG_ON(ret);
 425
 426         trace_and_count(c, btree_node_alloc, trans, b);
 427         bch2_increment_clock(c, btree_sectors(c), WRITE);
 428         return b;
 429 }
 430
 431 static void btree_set_min(struct btree *b, struct bpos pos)
 432 {
 433         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 434                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 435         b->data->min_key = pos;
 436 }
 437
 438 static void btree_set_max(struct btree *b, struct bpos pos)
 439 {
 440         b->key.k.p = pos;
 441         b->data->max_key = pos;
 442 }
 443
 444 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 445                                                        struct btree_trans *trans,
 446                                                        struct btree *b)
 447 {
 448         struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
 449         struct bkey_format format = bch2_btree_calc_format(b);
 450
 451         /*
 452          * The keys might expand with the new format - if they wouldn't fit in
 453          * the btree node anymore, use the old format for now:
 454          */
 455         if (!bch2_btree_node_format_fits(as->c, b, b->nr, &format))
 456                 format = b->format;
 457
 458         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 459
 460         btree_set_min(n, b->data->min_key);
 461         btree_set_max(n, b->data->max_key);
 462
 463         n->data->format         = format;
 464         btree_node_set_format(n, format);
 465
 466         bch2_btree_sort_into(as->c, n, b);
 467
 468         btree_node_reset_sib_u64s(n);
 469         return n;
 470 }
 471
 472 static struct btree *__btree_root_alloc(struct btree_update *as,
 473                                 struct btree_trans *trans, unsigned level)
 474 {
 475         struct btree *b = bch2_btree_node_alloc(as, trans, level);
 476
 477         btree_set_min(b, POS_MIN);
 478         btree_set_max(b, SPOS_MAX);
 479         b->data->format = bch2_btree_calc_format(b);
 480
 481         btree_node_set_format(b, b->data->format);
 482         bch2_btree_build_aux_trees(b);
 483
 484         return b;
 485 }
 486
 487 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
 488 {
 489         struct bch_fs *c = as->c;
 490         struct prealloc_nodes *p;
 491
 492         for (p = as->prealloc_nodes;
 493              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 494              p++) {
 495                 while (p->nr) {
 496                         struct btree *b = p->b[--p->nr];
 497
 498                         mutex_lock(&c->btree_reserve_cache_lock);
 499
 500                         if (c->btree_reserve_cache_nr <
 501                             ARRAY_SIZE(c->btree_reserve_cache)) {
 502                                 struct btree_alloc *a =
 503                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 504
 505                                 a->ob = b->ob;
 506                                 b->ob.nr = 0;
 507                                 bkey_copy(&a->k, &b->key);
 508                         } else {
 509                                 bch2_open_buckets_put(c, &b->ob);
 510                         }
 511
 512                         mutex_unlock(&c->btree_reserve_cache_lock);
 513
 514                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 515                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 516                         __btree_node_free(trans, b);
 517                         six_unlock_write(&b->c.lock);
 518                         six_unlock_intent(&b->c.lock);
 519                 }
 520         }
 521 }
 522
 523 static int bch2_btree_reserve_get(struct btree_trans *trans,
 524                                   struct btree_update *as,
 525                                   unsigned nr_nodes[2],
 526                                   unsigned flags,
 527                                   struct closure *cl)
 528 {
 529         struct btree *b;
 530         unsigned interior;
 531         int ret = 0;
 532
 533         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 534
 535         /*
 536          * Protects reaping from the btree node cache and using the btree node
 537          * open bucket reserve:
 538          */
 539         ret = bch2_btree_cache_cannibalize_lock(trans, cl);
 540         if (ret)
 541                 return ret;
 542
 543         for (interior = 0; interior < 2; interior++) {
 544                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 545
 546                 while (p->nr < nr_nodes[interior]) {
 547                         b = __bch2_btree_node_alloc(trans, &as->disk_res, cl,
 548                                                     interior, flags);
 549                         if (IS_ERR(b)) {
 550                                 ret = PTR_ERR(b);
 551                                 goto err;
 552                         }
 553
 554                         p->b[p->nr++] = b;
 555                 }
 556         }
 557 err:
 558         bch2_btree_cache_cannibalize_unlock(trans);
 559         return ret;
 560 }
 561
 562 /* Asynchronous interior node update machinery */
 563
 564 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
 565 {
 566         struct bch_fs *c = as->c;
 567
 568         if (as->took_gc_lock)
 569                 up_read(&c->gc_lock);
 570         as->took_gc_lock = false;
 571
 572         bch2_journal_pin_drop(&c->journal, &as->journal);
 573         bch2_journal_pin_flush(&c->journal, &as->journal);
 574         bch2_disk_reservation_put(c, &as->disk_res);
 575         bch2_btree_reserve_put(as, trans);
 576
 577         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 578                                as->start_time);
 579
 580         mutex_lock(&c->btree_interior_update_lock);
 581         list_del(&as->unwritten_list);
 582         list_del(&as->list);
 583
 584         closure_debug_destroy(&as->cl);
 585         mempool_free(as, &c->btree_interior_update_pool);
 586
 587         /*
 588          * Have to do the wakeup with btree_interior_update_lock still held,
 589          * since being on btree_interior_update_list is our ref on @c:
 590          */
 591         closure_wake_up(&c->btree_interior_update_wait);
 592
 593         mutex_unlock(&c->btree_interior_update_lock);
 594 }
 595
 596 static void btree_update_add_key(struct btree_update *as,
 597                                  struct keylist *keys, struct btree *b)
 598 {
 599         struct bkey_i *k = &b->key;
 600
 601         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 602                ARRAY_SIZE(as->_old_keys));
 603
 604         bkey_copy(keys->top, k);
 605         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 606
 607         bch2_keylist_push(keys);
 608 }
 609
 610 static bool btree_update_new_nodes_marked_sb(struct btree_update *as)
 611 {
 612         for_each_keylist_key(&as->new_keys, k)
 613                 if (!bch2_dev_btree_bitmap_marked(as->c, bkey_i_to_s_c(k)))
 614                         return false;
 615         return true;
 616 }
 617
 618 static void btree_update_new_nodes_mark_sb(struct btree_update *as)
 619 {
 620         struct bch_fs *c = as->c;
 621
 622         mutex_lock(&c->sb_lock);
 623         for_each_keylist_key(&as->new_keys, k)
 624                 bch2_dev_btree_bitmap_mark(c, bkey_i_to_s_c(k));
 625
 626         bch2_write_super(c);
 627         mutex_unlock(&c->sb_lock);
 628 }
 629
 630 /*
 631  * The transactional part of an interior btree node update, where we journal the
 632  * update we did to the interior node and update alloc info:
 633  */
 634 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 635                                             struct btree_update *as)
 636 {
 637         struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, as->journal_u64s);
 638         int ret = PTR_ERR_OR_ZERO(e);
 639         if (ret)
 640                 return ret;
 641
 642         memcpy(e, as->journal_entries, as->journal_u64s * sizeof(u64));
 643
 644         trans->journal_pin = &as->journal;
 645
 646         for_each_keylist_key(&as->old_keys, k) {
 647                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 648
 649                 ret = bch2_key_trigger_old(trans, as->btree_id, level, bkey_i_to_s_c(k),
 650                                            BTREE_TRIGGER_transactional);
 651                 if (ret)
 652                         return ret;
 653         }
 654
 655         for_each_keylist_key(&as->new_keys, k) {
 656                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 657
 658                 ret = bch2_key_trigger_new(trans, as->btree_id, level, bkey_i_to_s(k),
 659                                            BTREE_TRIGGER_transactional);
 660                 if (ret)
 661                         return ret;
 662         }
 663
 664         return 0;
 665 }
 666
 667 static void btree_update_nodes_written(struct btree_update *as)
 668 {
 669         struct bch_fs *c = as->c;
 670         struct btree *b;
 671         struct btree_trans *trans = bch2_trans_get(c);
 672         u64 journal_seq = 0;
 673         unsigned i;
 674         int ret;
 675
 676         /*
 677          * If we're already in an error state, it might be because a btree node
 678          * was never written, and we might be trying to free that same btree
 679          * node here, but it won't have been marked as allocated and we'll see
 680          * spurious disk usage inconsistencies in the transactional part below
 681          * if we don't skip it:
 682          */
 683         ret = bch2_journal_error(&c->journal);
 684         if (ret)
 685                 goto err;
 686
 687         if (!btree_update_new_nodes_marked_sb(as))
 688                 btree_update_new_nodes_mark_sb(as);
 689
 690         /*
 691          * Wait for any in flight writes to finish before we free the old nodes
 692          * on disk:
 693          */
 694         for (i = 0; i < as->nr_old_nodes; i++) {
 695                 __le64 seq;
 696
 697                 b = as->old_nodes[i];
 698
 699                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 700                 seq = b->data ? b->data->keys.seq : 0;
 701                 six_unlock_read(&b->c.lock);
 702
 703                 if (seq == as->old_nodes_seq[i])
 704                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 705                                        TASK_UNINTERRUPTIBLE);
 706         }
 707
 708         /*
 709          * We did an update to a parent node where the pointers we added pointed
 710          * to child nodes that weren't written yet: now, the child nodes have
 711          * been written so we can write out the update to the interior node.
 712          */
 713
 714         /*
 715          * We can't call into journal reclaim here: we'd block on the journal
 716          * reclaim lock, but we may need to release the open buckets we have
 717          * pinned in order for other btree updates to make forward progress, and
 718          * journal reclaim does btree updates when flushing bkey_cached entries,
 719          * which may require allocations as well.
 720          */
 721         ret = commit_do(trans, &as->disk_res, &journal_seq,
 722                         BCH_WATERMARK_interior_updates|
 723                         BCH_TRANS_COMMIT_no_enospc|
 724                         BCH_TRANS_COMMIT_no_check_rw|
 725                         BCH_TRANS_COMMIT_journal_reclaim,
 726                         btree_update_nodes_written_trans(trans, as));
 727         bch2_trans_unlock(trans);
 728
 729         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 730                              "%s", bch2_err_str(ret));
 731 err:
 732         /*
 733          * We have to be careful because another thread might be getting ready
 734          * to free as->b and calling btree_update_reparent() on us - we'll
 735          * recheck under btree_update_lock below:
 736          */
 737         b = READ_ONCE(as->b);
 738         if (b) {
 739                 /*
 740                  * @b is the node we did the final insert into:
 741                  *
 742                  * On failure to get a journal reservation, we still have to
 743                  * unblock the write and allow most of the write path to happen
 744                  * so that shutdown works, but the i->journal_seq mechanism
 745                  * won't work to prevent the btree write from being visible (we
 746                  * didn't get a journal sequence number) - instead
 747                  * __bch2_btree_node_write() doesn't do the actual write if
 748                  * we're in journal error state:
 749                  */
 750
 751                 /*
 752                  * Ensure transaction is unlocked before using
 753                  * btree_node_lock_nopath() (the use of which is always suspect,
 754                  * we need to work on removing this in the future)
 755                  *
 756                  * It should be, but bch2_path_get_unlocked_mut() -> bch2_path_get()
 757                  * calls bch2_path_upgrade(), before we call path_make_mut(), so
 758                  * we may rarely end up with a locked path besides the one we
 759                  * have here:
 760                  */
 761                 bch2_trans_unlock(trans);
 762                 bch2_trans_begin(trans);
 763                 btree_path_idx_t path_idx = bch2_path_get_unlocked_mut(trans,
 764                                                 as->btree_id, b->c.level, b->key.k.p);
 765                 struct btree_path *path = trans->paths + path_idx;
 766                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 767                 mark_btree_node_locked(trans, path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 768                 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 769                 path->l[b->c.level].b = b;
 770
 771                 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 772
 773                 mutex_lock(&c->btree_interior_update_lock);
 774
 775                 list_del(&as->write_blocked_list);
 776                 if (list_empty(&b->write_blocked))
 777                         clear_btree_node_write_blocked(b);
 778
 779                 /*
 780                  * Node might have been freed, recheck under
 781                  * btree_interior_update_lock:
 782                  */
 783                 if (as->b == b) {
 784                         BUG_ON(!b->c.level);
 785                         BUG_ON(!btree_node_dirty(b));
 786
 787                         if (!ret) {
 788                                 struct bset *last = btree_bset_last(b);
 789
 790                                 last->journal_seq = cpu_to_le64(
 791                                                              max(journal_seq,
 792                                                                  le64_to_cpu(last->journal_seq)));
 793
 794                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 795                         } else {
 796                                 /*
 797                                  * If we didn't get a journal sequence number we
 798                                  * can't write this btree node, because recovery
 799                                  * won't know to ignore this write:
 800                                  */
 801                                 set_btree_node_never_write(b);
 802                         }
 803                 }
 804
 805                 mutex_unlock(&c->btree_interior_update_lock);
 806
 807                 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 808                 six_unlock_write(&b->c.lock);
 809
 810                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 811                 btree_node_unlock(trans, path, b->c.level);
 812                 bch2_path_put(trans, path_idx, true);
 813         }
 814
 815         bch2_journal_pin_drop(&c->journal, &as->journal);
 816
 817         mutex_lock(&c->btree_interior_update_lock);
 818         for (i = 0; i < as->nr_new_nodes; i++) {
 819                 b = as->new_nodes[i];
 820
 821                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 822                 b->will_make_reachable = 0;
 823                 clear_btree_node_will_make_reachable(b);
 824         }
 825         mutex_unlock(&c->btree_interior_update_lock);
 826
 827         for (i = 0; i < as->nr_new_nodes; i++) {
 828                 b = as->new_nodes[i];
 829
 830                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 831                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 832                 six_unlock_read(&b->c.lock);
 833         }
 834
 835         for (i = 0; i < as->nr_open_buckets; i++)
 836                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 837
 838         bch2_btree_update_free(as, trans);
 839         bch2_trans_put(trans);
 840 }
 841
 842 static void btree_interior_update_work(struct work_struct *work)
 843 {
 844         struct bch_fs *c =
 845                 container_of(work, struct bch_fs, btree_interior_update_work);
 846         struct btree_update *as;
 847
 848         while (1) {
 849                 mutex_lock(&c->btree_interior_update_lock);
 850                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 851                                               struct btree_update, unwritten_list);
 852                 if (as && !as->nodes_written)
 853                         as = NULL;
 854                 mutex_unlock(&c->btree_interior_update_lock);
 855
 856                 if (!as)
 857                         break;
 858
 859                 btree_update_nodes_written(as);
 860         }
 861 }
 862
 863 static CLOSURE_CALLBACK(btree_update_set_nodes_written)
 864 {
 865         closure_type(as, struct btree_update, cl);
 866         struct bch_fs *c = as->c;
 867
 868         mutex_lock(&c->btree_interior_update_lock);
 869         as->nodes_written = true;
 870         mutex_unlock(&c->btree_interior_update_lock);
 871
 872         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 873 }
 874
 875 /*
 876  * We're updating @b with pointers to nodes that haven't finished writing yet:
 877  * block @b from being written until @as completes
 878  */
 879 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 880 {
 881         struct bch_fs *c = as->c;
 882
 883         BUG_ON(as->mode != BTREE_UPDATE_none);
 884         BUG_ON(as->update_level_end < b->c.level);
 885         BUG_ON(!btree_node_dirty(b));
 886         BUG_ON(!b->c.level);
 887
 888         mutex_lock(&c->btree_interior_update_lock);
 889         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 890
 891         as->mode        = BTREE_UPDATE_node;
 892         as->b           = b;
 893         as->update_level_end = b->c.level;
 894
 895         set_btree_node_write_blocked(b);
 896         list_add(&as->write_blocked_list, &b->write_blocked);
 897
 898         mutex_unlock(&c->btree_interior_update_lock);
 899 }
 900
 901 static int bch2_update_reparent_journal_pin_flush(struct journal *j,
 902                                 struct journal_entry_pin *_pin, u64 seq)
 903 {
 904         return 0;
 905 }
 906
 907 static void btree_update_reparent(struct btree_update *as,
 908                                   struct btree_update *child)
 909 {
 910         struct bch_fs *c = as->c;
 911
 912         lockdep_assert_held(&c->btree_interior_update_lock);
 913
 914         child->b = NULL;
 915         child->mode = BTREE_UPDATE_update;
 916
 917         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal,
 918                               bch2_update_reparent_journal_pin_flush);
 919 }
 920
 921 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 922 {
 923         struct bkey_i *insert = &b->key;
 924         struct bch_fs *c = as->c;
 925
 926         BUG_ON(as->mode != BTREE_UPDATE_none);
 927
 928         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 929                ARRAY_SIZE(as->journal_entries));
 930
 931         as->journal_u64s +=
 932                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 933                                   BCH_JSET_ENTRY_btree_root,
 934                                   b->c.btree_id, b->c.level,
 935                                   insert, insert->k.u64s);
 936
 937         mutex_lock(&c->btree_interior_update_lock);
 938         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 939
 940         as->mode        = BTREE_UPDATE_root;
 941         mutex_unlock(&c->btree_interior_update_lock);
 942 }
 943
 944 /*
 945  * bch2_btree_update_add_new_node:
 946  *
 947  * This causes @as to wait on @b to be written, before it gets to
 948  * bch2_btree_update_nodes_written
 949  *
 950  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 951  * to @b from happening besides the first until @b is reachable on disk
 952  *
 953  * And it adds @b to the list of @as's new nodes, so that we can update sector
 954  * counts in bch2_btree_update_nodes_written:
 955  */
 956 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 957 {
 958         struct bch_fs *c = as->c;
 959
 960         closure_get(&as->cl);
 961
 962         mutex_lock(&c->btree_interior_update_lock);
 963         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 964         BUG_ON(b->will_make_reachable);
 965
 966         as->new_nodes[as->nr_new_nodes++] = b;
 967         b->will_make_reachable = 1UL|(unsigned long) as;
 968         set_btree_node_will_make_reachable(b);
 969
 970         mutex_unlock(&c->btree_interior_update_lock);
 971
 972         btree_update_add_key(as, &as->new_keys, b);
 973
 974         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 975                 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
 976                 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
 977
 978                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
 979                         cpu_to_le16(sectors);
 980         }
 981 }
 982
 983 /*
 984  * returns true if @b was a new node
 985  */
 986 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 987 {
 988         struct btree_update *as;
 989         unsigned long v;
 990         unsigned i;
 991
 992         mutex_lock(&c->btree_interior_update_lock);
 993         /*
 994          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 995          * dropped when it gets written by bch2_btree_complete_write - the
 996          * xchg() is for synchronization with bch2_btree_complete_write:
 997          */
 998         v = xchg(&b->will_make_reachable, 0);
 999         clear_btree_node_will_make_reachable(b);
1000         as = (struct btree_update *) (v & ~1UL);
1001
1002         if (!as) {
1003                 mutex_unlock(&c->btree_interior_update_lock);
1004                 return;
1005         }
1006
1007         for (i = 0; i < as->nr_new_nodes; i++)
1008                 if (as->new_nodes[i] == b)
1009                         goto found;
1010
1011         BUG();
1012 found:
1013         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
1014         mutex_unlock(&c->btree_interior_update_lock);
1015
1016         if (v & 1)
1017                 closure_put(&as->cl);
1018 }
1019
1020 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
1021 {
1022         while (b->ob.nr)
1023                 as->open_buckets[as->nr_open_buckets++] =
1024                         b->ob.v[--b->ob.nr];
1025 }
1026
1027 static int bch2_btree_update_will_free_node_journal_pin_flush(struct journal *j,
1028                                 struct journal_entry_pin *_pin, u64 seq)
1029 {
1030         return 0;
1031 }
1032
1033 /*
1034  * @b is being split/rewritten: it may have pointers to not-yet-written btree
1035  * nodes and thus outstanding btree_updates - redirect @b's
1036  * btree_updates to point to this btree_update:
1037  */
1038 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
1039                                                       struct btree *b)
1040 {
1041         struct bch_fs *c = as->c;
1042         struct btree_update *p, *n;
1043         struct btree_write *w;
1044
1045         set_btree_node_dying(b);
1046
1047         if (btree_node_fake(b))
1048                 return;
1049
1050         mutex_lock(&c->btree_interior_update_lock);
1051
1052         /*
1053          * Does this node have any btree_update operations preventing
1054          * it from being written?
1055          *
1056          * If so, redirect them to point to this btree_update: we can
1057          * write out our new nodes, but we won't make them visible until those
1058          * operations complete
1059          */
1060         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
1061                 list_del_init(&p->write_blocked_list);
1062                 btree_update_reparent(as, p);
1063
1064                 /*
1065                  * for flush_held_btree_writes() waiting on updates to flush or
1066                  * nodes to be writeable:
1067                  */
1068                 closure_wake_up(&c->btree_interior_update_wait);
1069         }
1070
1071         clear_btree_node_dirty_acct(c, b);
1072         clear_btree_node_need_write(b);
1073         clear_btree_node_write_blocked(b);
1074
1075         /*
1076          * Does this node have unwritten data that has a pin on the journal?
1077          *
1078          * If so, transfer that pin to the btree_update operation -
1079          * note that if we're freeing multiple nodes, we only need to keep the
1080          * oldest pin of any of the nodes we're freeing. We'll release the pin
1081          * when the new nodes are persistent and reachable on disk:
1082          */
1083         w = btree_current_write(b);
1084         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1085                               bch2_btree_update_will_free_node_journal_pin_flush);
1086         bch2_journal_pin_drop(&c->journal, &w->journal);
1087
1088         w = btree_prev_write(b);
1089         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1090                               bch2_btree_update_will_free_node_journal_pin_flush);
1091         bch2_journal_pin_drop(&c->journal, &w->journal);
1092
1093         mutex_unlock(&c->btree_interior_update_lock);
1094
1095         /*
1096          * Is this a node that isn't reachable on disk yet?
1097          *
1098          * Nodes that aren't reachable yet have writes blocked until they're
1099          * reachable - now that we've cancelled any pending writes and moved
1100          * things waiting on that write to wait on this update, we can drop this
1101          * node from the list of nodes that the other update is making
1102          * reachable, prior to freeing it:
1103          */
1104         btree_update_drop_new_node(c, b);
1105
1106         btree_update_add_key(as, &as->old_keys, b);
1107
1108         as->old_nodes[as->nr_old_nodes] = b;
1109         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1110         as->nr_old_nodes++;
1111 }
1112
1113 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1114 {
1115         struct bch_fs *c = as->c;
1116         u64 start_time = as->start_time;
1117
1118         BUG_ON(as->mode == BTREE_UPDATE_none);
1119
1120         if (as->took_gc_lock)
1121                 up_read(&as->c->gc_lock);
1122         as->took_gc_lock = false;
1123
1124         bch2_btree_reserve_put(as, trans);
1125
1126         continue_at(&as->cl, btree_update_set_nodes_written,
1127                     as->c->btree_interior_update_worker);
1128
1129         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1130                                start_time);
1131 }
1132
1133 static struct btree_update *
1134 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1135                         unsigned level_start, bool split, unsigned flags)
1136 {
1137         struct bch_fs *c = trans->c;
1138         struct btree_update *as;
1139         u64 start_time = local_clock();
1140         int disk_res_flags = (flags & BCH_TRANS_COMMIT_no_enospc)
1141                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1142         unsigned nr_nodes[2] = { 0, 0 };
1143         unsigned level_end = level_start;
1144         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1145         int ret = 0;
1146         u32 restart_count = trans->restart_count;
1147
1148         BUG_ON(!path->should_be_locked);
1149
1150         if (watermark == BCH_WATERMARK_copygc)
1151                 watermark = BCH_WATERMARK_btree_copygc;
1152         if (watermark < BCH_WATERMARK_btree)
1153                 watermark = BCH_WATERMARK_btree;
1154
1155         flags &= ~BCH_WATERMARK_MASK;
1156         flags |= watermark;
1157
1158         if (watermark < BCH_WATERMARK_reclaim &&
1159             test_bit(JOURNAL_space_low, &c->journal.flags)) {
1160                 if (flags & BCH_TRANS_COMMIT_journal_reclaim)
1161                         return ERR_PTR(-BCH_ERR_journal_reclaim_would_deadlock);
1162
1163                 ret = drop_locks_do(trans,
1164                         ({ wait_event(c->journal.wait, !test_bit(JOURNAL_space_low, &c->journal.flags)); 0; }));
1165                 if (ret)
1166                         return ERR_PTR(ret);
1167         }
1168
1169         while (1) {
1170                 nr_nodes[!!level_end] += 1 + split;
1171                 level_end++;
1172
1173                 ret = bch2_btree_path_upgrade(trans, path, level_end + 1);
1174                 if (ret)
1175                         return ERR_PTR(ret);
1176
1177                 if (!btree_path_node(path, level_end)) {
1178                         /* Allocating new root? */
1179                         nr_nodes[1] += split;
1180                         level_end = BTREE_MAX_DEPTH;
1181                         break;
1182                 }
1183
1184                 /*
1185                  * Always check for space for two keys, even if we won't have to
1186                  * split at prior level - it might have been a merge instead:
1187                  */
1188                 if (bch2_btree_node_insert_fits(path->l[level_end].b,
1189                                                 BKEY_BTREE_PTR_U64s_MAX * 2))
1190                         break;
1191
1192                 split = path->l[level_end].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1193         }
1194
1195         if (!down_read_trylock(&c->gc_lock)) {
1196                 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1197                 if (ret) {
1198                         up_read(&c->gc_lock);
1199                         return ERR_PTR(ret);
1200                 }
1201         }
1202
1203         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1204         memset(as, 0, sizeof(*as));
1205         closure_init(&as->cl, NULL);
1206         as->c                   = c;
1207         as->start_time          = start_time;
1208         as->ip_started          = _RET_IP_;
1209         as->mode                = BTREE_UPDATE_none;
1210         as->flags               = flags;
1211         as->took_gc_lock        = true;
1212         as->btree_id            = path->btree_id;
1213         as->update_level_start  = level_start;
1214         as->update_level_end    = level_end;
1215         INIT_LIST_HEAD(&as->list);
1216         INIT_LIST_HEAD(&as->unwritten_list);
1217         INIT_LIST_HEAD(&as->write_blocked_list);
1218         bch2_keylist_init(&as->old_keys, as->_old_keys);
1219         bch2_keylist_init(&as->new_keys, as->_new_keys);
1220         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1221
1222         mutex_lock(&c->btree_interior_update_lock);
1223         list_add_tail(&as->list, &c->btree_interior_update_list);
1224         mutex_unlock(&c->btree_interior_update_lock);
1225
1226         /*
1227          * We don't want to allocate if we're in an error state, that can cause
1228          * deadlock on emergency shutdown due to open buckets getting stuck in
1229          * the btree_reserve_cache after allocator shutdown has cleared it out.
1230          * This check needs to come after adding us to the btree_interior_update
1231          * list but before calling bch2_btree_reserve_get, to synchronize with
1232          * __bch2_fs_read_only().
1233          */
1234         ret = bch2_journal_error(&c->journal);
1235         if (ret)
1236                 goto err;
1237
1238         ret = bch2_disk_reservation_get(c, &as->disk_res,
1239                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1240                         c->opts.metadata_replicas,
1241                         disk_res_flags);
1242         if (ret)
1243                 goto err;
1244
1245         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1246         if (bch2_err_matches(ret, ENOSPC) ||
1247             bch2_err_matches(ret, ENOMEM)) {
1248                 struct closure cl;
1249
1250                 /*
1251                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1252                  * flag
1253                  */
1254                 if (bch2_err_matches(ret, ENOSPC) &&
1255                     (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1256                     watermark < BCH_WATERMARK_reclaim) {
1257                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1258                         goto err;
1259                 }
1260
1261                 closure_init_stack(&cl);
1262
1263                 do {
1264                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1265
1266                         bch2_trans_unlock(trans);
1267                         closure_sync(&cl);
1268                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1269         }
1270
1271         if (ret) {
1272                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1273                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1274                 goto err;
1275         }
1276
1277         ret = bch2_trans_relock(trans);
1278         if (ret)
1279                 goto err;
1280
1281         bch2_trans_verify_not_restarted(trans, restart_count);
1282         return as;
1283 err:
1284         bch2_btree_update_free(as, trans);
1285         if (!bch2_err_matches(ret, ENOSPC) &&
1286             !bch2_err_matches(ret, EROFS) &&
1287             ret != -BCH_ERR_journal_reclaim_would_deadlock)
1288                 bch_err_fn_ratelimited(c, ret);
1289         return ERR_PTR(ret);
1290 }
1291
1292 /* Btree root updates: */
1293
1294 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1295 {
1296         /* Root nodes cannot be reaped */
1297         mutex_lock(&c->btree_cache.lock);
1298         list_del_init(&b->list);
1299         mutex_unlock(&c->btree_cache.lock);
1300
1301         mutex_lock(&c->btree_root_lock);
1302         bch2_btree_id_root(c, b->c.btree_id)->b = b;
1303         mutex_unlock(&c->btree_root_lock);
1304
1305         bch2_recalc_btree_reserve(c);
1306 }
1307
1308 static int bch2_btree_set_root(struct btree_update *as,
1309                                struct btree_trans *trans,
1310                                struct btree_path *path,
1311                                struct btree *b,
1312                                bool nofail)
1313 {
1314         struct bch_fs *c = as->c;
1315
1316         trace_and_count(c, btree_node_set_root, trans, b);
1317
1318         struct btree *old = btree_node_root(c, b);
1319
1320         /*
1321          * Ensure no one is using the old root while we switch to the
1322          * new root:
1323          */
1324         if (nofail) {
1325                 bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1326         } else {
1327                 int ret = bch2_btree_node_lock_write(trans, path, &old->c);
1328                 if (ret)
1329                         return ret;
1330         }
1331
1332         bch2_btree_set_root_inmem(c, b);
1333
1334         btree_update_updated_root(as, b);
1335
1336         /*
1337          * Unlock old root after new root is visible:
1338          *
1339          * The new root isn't persistent, but that's ok: we still have
1340          * an intent lock on the new root, and any updates that would
1341          * depend on the new root would have to update the new root.
1342          */
1343         bch2_btree_node_unlock_write(trans, path, old);
1344         return 0;
1345 }
1346
1347 /* Interior node updates: */
1348
1349 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1350                                         struct btree_trans *trans,
1351                                         struct btree_path *path,
1352                                         struct btree *b,
1353                                         struct btree_node_iter *node_iter,
1354                                         struct bkey_i *insert)
1355 {
1356         struct bch_fs *c = as->c;
1357         struct bkey_packed *k;
1358         struct printbuf buf = PRINTBUF;
1359         unsigned long old, new, v;
1360
1361         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1362                !btree_ptr_sectors_written(insert));
1363
1364         if (unlikely(!test_bit(JOURNAL_replay_done, &c->journal.flags)))
1365                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1366
1367         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1368                               btree_node_type(b), WRITE, &buf) ?:
1369             bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) {
1370                 printbuf_reset(&buf);
1371                 prt_printf(&buf, "inserting invalid bkey\n  ");
1372                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1373                 prt_printf(&buf, "\n  ");
1374                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1375                                   btree_node_type(b), WRITE, &buf);
1376                 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf);
1377
1378                 bch2_fs_inconsistent(c, "%s", buf.buf);
1379                 dump_stack();
1380         }
1381
1382         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1383                ARRAY_SIZE(as->journal_entries));
1384
1385         as->journal_u64s +=
1386                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1387                                   BCH_JSET_ENTRY_btree_keys,
1388                                   b->c.btree_id, b->c.level,
1389                                   insert, insert->k.u64s);
1390
1391         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1392                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1393                 bch2_btree_node_iter_advance(node_iter, b);
1394
1395         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1396         set_btree_node_dirty_acct(c, b);
1397
1398         v = READ_ONCE(b->flags);
1399         do {
1400                 old = new = v;
1401
1402                 new &= ~BTREE_WRITE_TYPE_MASK;
1403                 new |= BTREE_WRITE_interior;
1404                 new |= 1 << BTREE_NODE_need_write;
1405         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1406
1407         printbuf_exit(&buf);
1408 }
1409
1410 static void
1411 bch2_btree_insert_keys_interior(struct btree_update *as,
1412                                 struct btree_trans *trans,
1413                                 struct btree_path *path,
1414                                 struct btree *b,
1415                                 struct btree_node_iter node_iter,
1416                                 struct keylist *keys)
1417 {
1418         struct bkey_i *insert = bch2_keylist_front(keys);
1419         struct bkey_packed *k;
1420
1421         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1422
1423         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1424                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1425                 ;
1426
1427         while (!bch2_keylist_empty(keys)) {
1428                 insert = bch2_keylist_front(keys);
1429
1430                 if (bpos_gt(insert->k.p, b->key.k.p))
1431                         break;
1432
1433                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1434                 bch2_keylist_pop_front(keys);
1435         }
1436 }
1437
1438 /*
1439  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1440  * node)
1441  */
1442 static void __btree_split_node(struct btree_update *as,
1443                                struct btree_trans *trans,
1444                                struct btree *b,
1445                                struct btree *n[2])
1446 {
1447         struct bkey_packed *k;
1448         struct bpos n1_pos = POS_MIN;
1449         struct btree_node_iter iter;
1450         struct bset *bsets[2];
1451         struct bkey_format_state format[2];
1452         struct bkey_packed *out[2];
1453         struct bkey uk;
1454         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1455         struct { unsigned nr_keys, val_u64s; } nr_keys[2];
1456         int i;
1457
1458         memset(&nr_keys, 0, sizeof(nr_keys));
1459
1460         for (i = 0; i < 2; i++) {
1461                 BUG_ON(n[i]->nsets != 1);
1462
1463                 bsets[i] = btree_bset_first(n[i]);
1464                 out[i] = bsets[i]->start;
1465
1466                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1467                 bch2_bkey_format_init(&format[i]);
1468         }
1469
1470         u64s = 0;
1471         for_each_btree_node_key(b, k, &iter) {
1472                 if (bkey_deleted(k))
1473                         continue;
1474
1475                 uk = bkey_unpack_key(b, k);
1476
1477                 if (b->c.level &&
1478                     u64s < n1_u64s &&
1479                     u64s + k->u64s >= n1_u64s &&
1480                     bch2_key_deleted_in_journal(trans, b->c.btree_id, b->c.level, uk.p))
1481                         n1_u64s += k->u64s;
1482
1483                 i = u64s >= n1_u64s;
1484                 u64s += k->u64s;
1485                 if (!i)
1486                         n1_pos = uk.p;
1487                 bch2_bkey_format_add_key(&format[i], &uk);
1488
1489                 nr_keys[i].nr_keys++;
1490                 nr_keys[i].val_u64s += bkeyp_val_u64s(&b->format, k);
1491         }
1492
1493         btree_set_min(n[0], b->data->min_key);
1494         btree_set_max(n[0], n1_pos);
1495         btree_set_min(n[1], bpos_successor(n1_pos));
1496         btree_set_max(n[1], b->data->max_key);
1497
1498         for (i = 0; i < 2; i++) {
1499                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1500                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1501
1502                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1503
1504                 unsigned u64s = nr_keys[i].nr_keys * n[i]->data->format.key_u64s +
1505                         nr_keys[i].val_u64s;
1506                 if (__vstruct_bytes(struct btree_node, u64s) > btree_buf_bytes(b))
1507                         n[i]->data->format = b->format;
1508
1509                 btree_node_set_format(n[i], n[i]->data->format);
1510         }
1511
1512         u64s = 0;
1513         for_each_btree_node_key(b, k, &iter) {
1514                 if (bkey_deleted(k))
1515                         continue;
1516
1517                 i = u64s >= n1_u64s;
1518                 u64s += k->u64s;
1519
1520                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1521                                         ? &b->format: &bch2_bkey_format_current, k))
1522                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1523                 else
1524                         bch2_bkey_unpack(b, (void *) out[i], k);
1525
1526                 out[i]->needs_whiteout = false;
1527
1528                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1529                 out[i] = bkey_p_next(out[i]);
1530         }
1531
1532         for (i = 0; i < 2; i++) {
1533                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1534
1535                 BUG_ON(!bsets[i]->u64s);
1536
1537                 set_btree_bset_end(n[i], n[i]->set);
1538
1539                 btree_node_reset_sib_u64s(n[i]);
1540
1541                 bch2_verify_btree_nr_keys(n[i]);
1542
1543                 BUG_ON(bch2_btree_node_check_topology(trans, n[i]));
1544         }
1545 }
1546
1547 /*
1548  * For updates to interior nodes, we've got to do the insert before we split
1549  * because the stuff we're inserting has to be inserted atomically. Post split,
1550  * the keys might have to go in different nodes and the split would no longer be
1551  * atomic.
1552  *
1553  * Worse, if the insert is from btree node coalescing, if we do the insert after
1554  * we do the split (and pick the pivot) - the pivot we pick might be between
1555  * nodes that were coalesced, and thus in the middle of a child node post
1556  * coalescing:
1557  */
1558 static void btree_split_insert_keys(struct btree_update *as,
1559                                     struct btree_trans *trans,
1560                                     btree_path_idx_t path_idx,
1561                                     struct btree *b,
1562                                     struct keylist *keys)
1563 {
1564         struct btree_path *path = trans->paths + path_idx;
1565
1566         if (!bch2_keylist_empty(keys) &&
1567             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1568                 struct btree_node_iter node_iter;
1569
1570                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1571
1572                 bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1573
1574                 BUG_ON(bch2_btree_node_check_topology(trans, b));
1575         }
1576 }
1577
1578 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1579                        btree_path_idx_t path, struct btree *b,
1580                        struct keylist *keys)
1581 {
1582         struct bch_fs *c = as->c;
1583         struct btree *parent = btree_node_parent(trans->paths + path, b);
1584         struct btree *n1, *n2 = NULL, *n3 = NULL;
1585         btree_path_idx_t path1 = 0, path2 = 0;
1586         u64 start_time = local_clock();
1587         int ret = 0;
1588
1589         bch2_verify_btree_nr_keys(b);
1590         BUG_ON(!parent && (b != btree_node_root(c, b)));
1591         BUG_ON(parent && !btree_node_intent_locked(trans->paths + path, b->c.level + 1));
1592
1593         ret = bch2_btree_node_check_topology(trans, b);
1594         if (ret)
1595                 return ret;
1596
1597         bch2_btree_interior_update_will_free_node(as, b);
1598
1599         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1600                 struct btree *n[2];
1601
1602                 trace_and_count(c, btree_node_split, trans, b);
1603
1604                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1605                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1606
1607                 __btree_split_node(as, trans, b, n);
1608
1609                 if (keys) {
1610                         btree_split_insert_keys(as, trans, path, n1, keys);
1611                         btree_split_insert_keys(as, trans, path, n2, keys);
1612                         BUG_ON(!bch2_keylist_empty(keys));
1613                 }
1614
1615                 bch2_btree_build_aux_trees(n2);
1616                 bch2_btree_build_aux_trees(n1);
1617
1618                 bch2_btree_update_add_new_node(as, n1);
1619                 bch2_btree_update_add_new_node(as, n2);
1620                 six_unlock_write(&n2->c.lock);
1621                 six_unlock_write(&n1->c.lock);
1622
1623                 path1 = bch2_path_get_unlocked_mut(trans, as->btree_id, n1->c.level, n1->key.k.p);
1624                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1625                 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1626                 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1627
1628                 path2 = bch2_path_get_unlocked_mut(trans, as->btree_id, n2->c.level, n2->key.k.p);
1629                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1630                 mark_btree_node_locked(trans, trans->paths + path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1631                 bch2_btree_path_level_init(trans, trans->paths + path2, n2);
1632
1633                 /*
1634                  * Note that on recursive parent_keys == keys, so we
1635                  * can't start adding new keys to parent_keys before emptying it
1636                  * out (which we did with btree_split_insert_keys() above)
1637                  */
1638                 bch2_keylist_add(&as->parent_keys, &n1->key);
1639                 bch2_keylist_add(&as->parent_keys, &n2->key);
1640
1641                 if (!parent) {
1642                         /* Depth increases, make a new root */
1643                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1644
1645                         bch2_btree_update_add_new_node(as, n3);
1646                         six_unlock_write(&n3->c.lock);
1647
1648                         trans->paths[path2].locks_want++;
1649                         BUG_ON(btree_node_locked(trans->paths + path2, n3->c.level));
1650                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1651                         mark_btree_node_locked(trans, trans->paths + path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1652                         bch2_btree_path_level_init(trans, trans->paths + path2, n3);
1653
1654                         n3->sib_u64s[0] = U16_MAX;
1655                         n3->sib_u64s[1] = U16_MAX;
1656
1657                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1658                 }
1659         } else {
1660                 trace_and_count(c, btree_node_compact, trans, b);
1661
1662                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1663
1664                 if (keys) {
1665                         btree_split_insert_keys(as, trans, path, n1, keys);
1666                         BUG_ON(!bch2_keylist_empty(keys));
1667                 }
1668
1669                 bch2_btree_build_aux_trees(n1);
1670                 bch2_btree_update_add_new_node(as, n1);
1671                 six_unlock_write(&n1->c.lock);
1672
1673                 path1 = bch2_path_get_unlocked_mut(trans, as->btree_id, n1->c.level, n1->key.k.p);
1674                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1675                 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1676                 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1677
1678                 if (parent)
1679                         bch2_keylist_add(&as->parent_keys, &n1->key);
1680         }
1681
1682         /* New nodes all written, now make them visible: */
1683
1684         if (parent) {
1685                 /* Split a non root node */
1686                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys);
1687         } else if (n3) {
1688                 ret = bch2_btree_set_root(as, trans, trans->paths + path, n3, false);
1689         } else {
1690                 /* Root filled up but didn't need to be split */
1691                 ret = bch2_btree_set_root(as, trans, trans->paths + path, n1, false);
1692         }
1693
1694         if (ret)
1695                 goto err;
1696
1697         if (n3) {
1698                 bch2_btree_update_get_open_buckets(as, n3);
1699                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1700         }
1701         if (n2) {
1702                 bch2_btree_update_get_open_buckets(as, n2);
1703                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1704         }
1705         bch2_btree_update_get_open_buckets(as, n1);
1706         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1707
1708         /*
1709          * The old node must be freed (in memory) _before_ unlocking the new
1710          * nodes - else another thread could re-acquire a read lock on the old
1711          * node after another thread has locked and updated the new node, thus
1712          * seeing stale data:
1713          */
1714         bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1715
1716         if (n3)
1717                 bch2_trans_node_add(trans, trans->paths + path, n3);
1718         if (n2)
1719                 bch2_trans_node_add(trans, trans->paths + path2, n2);
1720         bch2_trans_node_add(trans, trans->paths + path1, n1);
1721
1722         if (n3)
1723                 six_unlock_intent(&n3->c.lock);
1724         if (n2)
1725                 six_unlock_intent(&n2->c.lock);
1726         six_unlock_intent(&n1->c.lock);
1727 out:
1728         if (path2) {
1729                 __bch2_btree_path_unlock(trans, trans->paths + path2);
1730                 bch2_path_put(trans, path2, true);
1731         }
1732         if (path1) {
1733                 __bch2_btree_path_unlock(trans, trans->paths + path1);
1734                 bch2_path_put(trans, path1, true);
1735         }
1736
1737         bch2_trans_verify_locks(trans);
1738
1739         bch2_time_stats_update(&c->times[n2
1740                                ? BCH_TIME_btree_node_split
1741                                : BCH_TIME_btree_node_compact],
1742                                start_time);
1743         return ret;
1744 err:
1745         if (n3)
1746                 bch2_btree_node_free_never_used(as, trans, n3);
1747         if (n2)
1748                 bch2_btree_node_free_never_used(as, trans, n2);
1749         bch2_btree_node_free_never_used(as, trans, n1);
1750         goto out;
1751 }
1752
1753 /**
1754  * bch2_btree_insert_node - insert bkeys into a given btree node
1755  *
1756  * @as:                 btree_update object
1757  * @trans:              btree_trans object
1758  * @path_idx:           path that points to current node
1759  * @b:                  node to insert keys into
1760  * @keys:               list of keys to insert
1761  *
1762  * Returns: 0 on success, typically transaction restart error on failure
1763  *
1764  * Inserts as many keys as it can into a given btree node, splitting it if full.
1765  * If a split occurred, this function will return early. This can only happen
1766  * for leaf nodes -- inserts into interior nodes have to be atomic.
1767  */
1768 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1769                                   btree_path_idx_t path_idx, struct btree *b,
1770                                   struct keylist *keys)
1771 {
1772         struct bch_fs *c = as->c;
1773         struct btree_path *path = trans->paths + path_idx, *linked;
1774         unsigned i;
1775         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1776         int old_live_u64s = b->nr.live_u64s;
1777         int live_u64s_added, u64s_added;
1778         int ret;
1779
1780         lockdep_assert_held(&c->gc_lock);
1781         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1782         BUG_ON(!b->c.level);
1783         BUG_ON(!as || as->b);
1784         bch2_verify_keylist_sorted(keys);
1785
1786         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1787         if (ret)
1788                 return ret;
1789
1790         bch2_btree_node_prep_for_write(trans, path, b);
1791
1792         if (!bch2_btree_node_insert_fits(b, bch2_keylist_u64s(keys))) {
1793                 bch2_btree_node_unlock_write(trans, path, b);
1794                 goto split;
1795         }
1796
1797         ret = bch2_btree_node_check_topology(trans, b);
1798         if (ret) {
1799                 bch2_btree_node_unlock_write(trans, path, b);
1800                 return ret;
1801         }
1802
1803         bch2_btree_insert_keys_interior(as, trans, path, b,
1804                                         path->l[b->c.level].iter, keys);
1805
1806         trans_for_each_path_with_node(trans, b, linked, i)
1807                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1808
1809         bch2_trans_verify_paths(trans);
1810
1811         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1812         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1813
1814         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1815                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1816         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1817                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1818
1819         if (u64s_added > live_u64s_added &&
1820             bch2_maybe_compact_whiteouts(c, b))
1821                 bch2_trans_node_reinit_iter(trans, b);
1822
1823         btree_update_updated_node(as, b);
1824         bch2_btree_node_unlock_write(trans, path, b);
1825
1826         BUG_ON(bch2_btree_node_check_topology(trans, b));
1827         return 0;
1828 split:
1829         /*
1830          * We could attempt to avoid the transaction restart, by calling
1831          * bch2_btree_path_upgrade() and allocating more nodes:
1832          */
1833         if (b->c.level >= as->update_level_end) {
1834                 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1835                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1836         }
1837
1838         return btree_split(as, trans, path_idx, b, keys);
1839 }
1840
1841 int bch2_btree_split_leaf(struct btree_trans *trans,
1842                           btree_path_idx_t path,
1843                           unsigned flags)
1844 {
1845         /* btree_split & merge may both cause paths array to be reallocated */
1846         struct btree *b = path_l(trans->paths + path)->b;
1847         struct btree_update *as;
1848         unsigned l;
1849         int ret = 0;
1850
1851         as = bch2_btree_update_start(trans, trans->paths + path,
1852                                      trans->paths[path].level,
1853                                      true, flags);
1854         if (IS_ERR(as))
1855                 return PTR_ERR(as);
1856
1857         ret = btree_split(as, trans, path, b, NULL);
1858         if (ret) {
1859                 bch2_btree_update_free(as, trans);
1860                 return ret;
1861         }
1862
1863         bch2_btree_update_done(as, trans);
1864
1865         for (l = trans->paths[path].level + 1;
1866              btree_node_intent_locked(&trans->paths[path], l) && !ret;
1867              l++)
1868                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1869
1870         return ret;
1871 }
1872
1873 static void __btree_increase_depth(struct btree_update *as, struct btree_trans *trans,
1874                                    btree_path_idx_t path_idx)
1875 {
1876         struct bch_fs *c = as->c;
1877         struct btree_path *path = trans->paths + path_idx;
1878         struct btree *n, *b = bch2_btree_id_root(c, path->btree_id)->b;
1879
1880         BUG_ON(!btree_node_locked(path, b->c.level));
1881
1882         n = __btree_root_alloc(as, trans, b->c.level + 1);
1883
1884         bch2_btree_update_add_new_node(as, n);
1885         six_unlock_write(&n->c.lock);
1886
1887         path->locks_want++;
1888         BUG_ON(btree_node_locked(path, n->c.level));
1889         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1890         mark_btree_node_locked(trans, path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1891         bch2_btree_path_level_init(trans, path, n);
1892
1893         n->sib_u64s[0] = U16_MAX;
1894         n->sib_u64s[1] = U16_MAX;
1895
1896         bch2_keylist_add(&as->parent_keys, &b->key);
1897         btree_split_insert_keys(as, trans, path_idx, n, &as->parent_keys);
1898
1899         int ret = bch2_btree_set_root(as, trans, path, n, true);
1900         BUG_ON(ret);
1901
1902         bch2_btree_update_get_open_buckets(as, n);
1903         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1904         bch2_trans_node_add(trans, path, n);
1905         six_unlock_intent(&n->c.lock);
1906
1907         mutex_lock(&c->btree_cache.lock);
1908         list_add_tail(&b->list, &c->btree_cache.live);
1909         mutex_unlock(&c->btree_cache.lock);
1910
1911         bch2_trans_verify_locks(trans);
1912 }
1913
1914 int bch2_btree_increase_depth(struct btree_trans *trans, btree_path_idx_t path, unsigned flags)
1915 {
1916         struct bch_fs *c = trans->c;
1917         struct btree *b = bch2_btree_id_root(c, trans->paths[path].btree_id)->b;
1918
1919         if (btree_node_fake(b))
1920                 return bch2_btree_split_leaf(trans, path, flags);
1921
1922         struct btree_update *as =
1923                 bch2_btree_update_start(trans, trans->paths + path, b->c.level, true, flags);
1924         if (IS_ERR(as))
1925                 return PTR_ERR(as);
1926
1927         __btree_increase_depth(as, trans, path);
1928         bch2_btree_update_done(as, trans);
1929         return 0;
1930 }
1931
1932 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1933                                   btree_path_idx_t path,
1934                                   unsigned level,
1935                                   unsigned flags,
1936                                   enum btree_node_sibling sib)
1937 {
1938         struct bch_fs *c = trans->c;
1939         struct btree_update *as;
1940         struct bkey_format_state new_s;
1941         struct bkey_format new_f;
1942         struct bkey_i delete;
1943         struct btree *b, *m, *n, *prev, *next, *parent;
1944         struct bpos sib_pos;
1945         size_t sib_u64s;
1946         enum btree_id btree = trans->paths[path].btree_id;
1947         btree_path_idx_t sib_path = 0, new_path = 0;
1948         u64 start_time = local_clock();
1949         int ret = 0;
1950
1951         bch2_trans_verify_not_in_restart(trans);
1952         bch2_trans_verify_not_unlocked(trans);
1953         BUG_ON(!trans->paths[path].should_be_locked);
1954         BUG_ON(!btree_node_locked(&trans->paths[path], level));
1955
1956         /*
1957          * Work around a deadlock caused by the btree write buffer not doing
1958          * merges and leaving tons of merges for us to do - we really don't need
1959          * to be doing merges at all from the interior update path, and if the
1960          * interior update path is generating too many new interior updates we
1961          * deadlock:
1962          */
1963         if ((flags & BCH_WATERMARK_MASK) == BCH_WATERMARK_interior_updates)
1964                 return 0;
1965
1966         if ((flags & BCH_WATERMARK_MASK) <= BCH_WATERMARK_reclaim) {
1967                 flags &= ~BCH_WATERMARK_MASK;
1968                 flags |= BCH_WATERMARK_btree;
1969                 flags |= BCH_TRANS_COMMIT_journal_reclaim;
1970         }
1971
1972         b = trans->paths[path].l[level].b;
1973
1974         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1975             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1976                 b->sib_u64s[sib] = U16_MAX;
1977                 return 0;
1978         }
1979
1980         sib_pos = sib == btree_prev_sib
1981                 ? bpos_predecessor(b->data->min_key)
1982                 : bpos_successor(b->data->max_key);
1983
1984         sib_path = bch2_path_get(trans, btree, sib_pos,
1985                                  U8_MAX, level, BTREE_ITER_intent, _THIS_IP_);
1986         ret = bch2_btree_path_traverse(trans, sib_path, false);
1987         if (ret)
1988                 goto err;
1989
1990         btree_path_set_should_be_locked(trans->paths + sib_path);
1991
1992         m = trans->paths[sib_path].l[level].b;
1993
1994         if (btree_node_parent(trans->paths + path, b) !=
1995             btree_node_parent(trans->paths + sib_path, m)) {
1996                 b->sib_u64s[sib] = U16_MAX;
1997                 goto out;
1998         }
1999
2000         if (sib == btree_prev_sib) {
2001                 prev = m;
2002                 next = b;
2003         } else {
2004                 prev = b;
2005                 next = m;
2006         }
2007
2008         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
2009                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
2010
2011                 bch2_bpos_to_text(&buf1, prev->data->max_key);
2012                 bch2_bpos_to_text(&buf2, next->data->min_key);
2013                 bch_err(c,
2014                         "%s(): btree topology error:\n"
2015                         "  prev ends at   %s\n"
2016                         "  next starts at %s",
2017                         __func__, buf1.buf, buf2.buf);
2018                 printbuf_exit(&buf1);
2019                 printbuf_exit(&buf2);
2020                 ret = bch2_topology_error(c);
2021                 goto err;
2022         }
2023
2024         bch2_bkey_format_init(&new_s);
2025         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
2026         __bch2_btree_calc_format(&new_s, prev);
2027         __bch2_btree_calc_format(&new_s, next);
2028         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
2029         new_f = bch2_bkey_format_done(&new_s);
2030
2031         sib_u64s = btree_node_u64s_with_format(b->nr, &b->format, &new_f) +
2032                 btree_node_u64s_with_format(m->nr, &m->format, &new_f);
2033
2034         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
2035                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
2036                 sib_u64s /= 2;
2037                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
2038         }
2039
2040         sib_u64s = min(sib_u64s, btree_max_u64s(c));
2041         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
2042         b->sib_u64s[sib] = sib_u64s;
2043
2044         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
2045                 goto out;
2046
2047         parent = btree_node_parent(trans->paths + path, b);
2048         as = bch2_btree_update_start(trans, trans->paths + path, level, false,
2049                                      BCH_TRANS_COMMIT_no_enospc|flags);
2050         ret = PTR_ERR_OR_ZERO(as);
2051         if (ret)
2052                 goto err;
2053
2054         trace_and_count(c, btree_node_merge, trans, b);
2055
2056         bch2_btree_interior_update_will_free_node(as, b);
2057         bch2_btree_interior_update_will_free_node(as, m);
2058
2059         n = bch2_btree_node_alloc(as, trans, b->c.level);
2060
2061         SET_BTREE_NODE_SEQ(n->data,
2062                            max(BTREE_NODE_SEQ(b->data),
2063                                BTREE_NODE_SEQ(m->data)) + 1);
2064
2065         btree_set_min(n, prev->data->min_key);
2066         btree_set_max(n, next->data->max_key);
2067
2068         n->data->format  = new_f;
2069         btree_node_set_format(n, new_f);
2070
2071         bch2_btree_sort_into(c, n, prev);
2072         bch2_btree_sort_into(c, n, next);
2073
2074         bch2_btree_build_aux_trees(n);
2075         bch2_btree_update_add_new_node(as, n);
2076         six_unlock_write(&n->c.lock);
2077
2078         new_path = bch2_path_get_unlocked_mut(trans, btree, n->c.level, n->key.k.p);
2079         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
2080         mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
2081         bch2_btree_path_level_init(trans, trans->paths + new_path, n);
2082
2083         bkey_init(&delete.k);
2084         delete.k.p = prev->key.k.p;
2085         bch2_keylist_add(&as->parent_keys, &delete);
2086         bch2_keylist_add(&as->parent_keys, &n->key);
2087
2088         bch2_trans_verify_paths(trans);
2089
2090         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys);
2091         if (ret)
2092                 goto err_free_update;
2093
2094         bch2_trans_verify_paths(trans);
2095
2096         bch2_btree_update_get_open_buckets(as, n);
2097         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
2098
2099         bch2_btree_node_free_inmem(trans, trans->paths + path, b);
2100         bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m);
2101
2102         bch2_trans_node_add(trans, trans->paths + path, n);
2103
2104         bch2_trans_verify_paths(trans);
2105
2106         six_unlock_intent(&n->c.lock);
2107
2108         bch2_btree_update_done(as, trans);
2109
2110         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
2111 out:
2112 err:
2113         if (new_path)
2114                 bch2_path_put(trans, new_path, true);
2115         bch2_path_put(trans, sib_path, true);
2116         bch2_trans_verify_locks(trans);
2117         if (ret == -BCH_ERR_journal_reclaim_would_deadlock)
2118                 ret = 0;
2119         if (!ret)
2120                 ret = bch2_trans_relock(trans);
2121         return ret;
2122 err_free_update:
2123         bch2_btree_node_free_never_used(as, trans, n);
2124         bch2_btree_update_free(as, trans);
2125         goto out;
2126 }
2127
2128 int bch2_btree_node_rewrite(struct btree_trans *trans,
2129                             struct btree_iter *iter,
2130                             struct btree *b,
2131                             unsigned flags)
2132 {
2133         struct bch_fs *c = trans->c;
2134         struct btree *n, *parent;
2135         struct btree_update *as;
2136         btree_path_idx_t new_path = 0;
2137         int ret;
2138
2139         flags |= BCH_TRANS_COMMIT_no_enospc;
2140
2141         struct btree_path *path = btree_iter_path(trans, iter);
2142         parent = btree_node_parent(path, b);
2143         as = bch2_btree_update_start(trans, path, b->c.level, false, flags);
2144         ret = PTR_ERR_OR_ZERO(as);
2145         if (ret)
2146                 goto out;
2147
2148         bch2_btree_interior_update_will_free_node(as, b);
2149
2150         n = bch2_btree_node_alloc_replacement(as, trans, b);
2151
2152         bch2_btree_build_aux_trees(n);
2153         bch2_btree_update_add_new_node(as, n);
2154         six_unlock_write(&n->c.lock);
2155
2156         new_path = bch2_path_get_unlocked_mut(trans, iter->btree_id, n->c.level, n->key.k.p);
2157         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
2158         mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
2159         bch2_btree_path_level_init(trans, trans->paths + new_path, n);
2160
2161         trace_and_count(c, btree_node_rewrite, trans, b);
2162
2163         if (parent) {
2164                 bch2_keylist_add(&as->parent_keys, &n->key);
2165                 ret = bch2_btree_insert_node(as, trans, iter->path, parent, &as->parent_keys);
2166         } else {
2167                 ret = bch2_btree_set_root(as, trans, btree_iter_path(trans, iter), n, false);
2168         }
2169
2170         if (ret)
2171                 goto err;
2172
2173         bch2_btree_update_get_open_buckets(as, n);
2174         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
2175
2176         bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b);
2177
2178         bch2_trans_node_add(trans, trans->paths + iter->path, n);
2179         six_unlock_intent(&n->c.lock);
2180
2181         bch2_btree_update_done(as, trans);
2182 out:
2183         if (new_path)
2184                 bch2_path_put(trans, new_path, true);
2185         bch2_trans_downgrade(trans);
2186         return ret;
2187 err:
2188         bch2_btree_node_free_never_used(as, trans, n);
2189         bch2_btree_update_free(as, trans);
2190         goto out;
2191 }
2192
2193 struct async_btree_rewrite {
2194         struct bch_fs           *c;
2195         struct work_struct      work;
2196         struct list_head        list;
2197         enum btree_id           btree_id;
2198         unsigned                level;
2199         struct bpos             pos;
2200         __le64                  seq;
2201 };
2202
2203 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2204                                           struct async_btree_rewrite *a)
2205 {
2206         struct bch_fs *c = trans->c;
2207         struct btree_iter iter;
2208         struct btree *b;
2209         int ret;
2210
2211         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2212                                   BTREE_MAX_DEPTH, a->level, 0);
2213         b = bch2_btree_iter_peek_node(&iter);
2214         ret = PTR_ERR_OR_ZERO(b);
2215         if (ret)
2216                 goto out;
2217
2218         if (!b || b->data->keys.seq != a->seq) {
2219                 struct printbuf buf = PRINTBUF;
2220
2221                 if (b)
2222                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2223                 else
2224                         prt_str(&buf, "(null");
2225                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2226                          __func__, a->seq, buf.buf);
2227                 printbuf_exit(&buf);
2228                 goto out;
2229         }
2230
2231         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2232 out:
2233         bch2_trans_iter_exit(trans, &iter);
2234
2235         return ret;
2236 }
2237
2238 static void async_btree_node_rewrite_work(struct work_struct *work)
2239 {
2240         struct async_btree_rewrite *a =
2241                 container_of(work, struct async_btree_rewrite, work);
2242         struct bch_fs *c = a->c;
2243         int ret;
2244
2245         ret = bch2_trans_do(c, NULL, NULL, 0,
2246                       async_btree_node_rewrite_trans(trans, a));
2247         bch_err_fn_ratelimited(c, ret);
2248         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2249         kfree(a);
2250 }
2251
2252 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2253 {
2254         struct async_btree_rewrite *a;
2255         int ret;
2256
2257         a = kmalloc(sizeof(*a), GFP_NOFS);
2258         if (!a) {
2259                 bch_err(c, "%s: error allocating memory", __func__);
2260                 return;
2261         }
2262
2263         a->c            = c;
2264         a->btree_id     = b->c.btree_id;
2265         a->level        = b->c.level;
2266         a->pos          = b->key.k.p;
2267         a->seq          = b->data->keys.seq;
2268         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2269
2270         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
2271                 mutex_lock(&c->pending_node_rewrites_lock);
2272                 list_add(&a->list, &c->pending_node_rewrites);
2273                 mutex_unlock(&c->pending_node_rewrites_lock);
2274                 return;
2275         }
2276
2277         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2278                 if (test_bit(BCH_FS_started, &c->flags)) {
2279                         bch_err(c, "%s: error getting c->writes ref", __func__);
2280                         kfree(a);
2281                         return;
2282                 }
2283
2284                 ret = bch2_fs_read_write_early(c);
2285                 bch_err_msg(c, ret, "going read-write");
2286                 if (ret) {
2287                         kfree(a);
2288                         return;
2289                 }
2290
2291                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2292         }
2293
2294         queue_work(c->btree_node_rewrite_worker, &a->work);
2295 }
2296
2297 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2298 {
2299         struct async_btree_rewrite *a, *n;
2300
2301         mutex_lock(&c->pending_node_rewrites_lock);
2302         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2303                 list_del(&a->list);
2304
2305                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2306                 queue_work(c->btree_node_rewrite_worker, &a->work);
2307         }
2308         mutex_unlock(&c->pending_node_rewrites_lock);
2309 }
2310
2311 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2312 {
2313         struct async_btree_rewrite *a, *n;
2314
2315         mutex_lock(&c->pending_node_rewrites_lock);
2316         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2317                 list_del(&a->list);
2318
2319                 kfree(a);
2320         }
2321         mutex_unlock(&c->pending_node_rewrites_lock);
2322 }
2323
2324 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2325                                         struct btree_iter *iter,
2326                                         struct btree *b, struct btree *new_hash,
2327                                         struct bkey_i *new_key,
2328                                         unsigned commit_flags,
2329                                         bool skip_triggers)
2330 {
2331         struct bch_fs *c = trans->c;
2332         struct btree_iter iter2 = { NULL };
2333         struct btree *parent;
2334         int ret;
2335
2336         if (!skip_triggers) {
2337                 ret   = bch2_key_trigger_old(trans, b->c.btree_id, b->c.level + 1,
2338                                              bkey_i_to_s_c(&b->key),
2339                                              BTREE_TRIGGER_transactional) ?:
2340                         bch2_key_trigger_new(trans, b->c.btree_id, b->c.level + 1,
2341                                              bkey_i_to_s(new_key),
2342                                              BTREE_TRIGGER_transactional);
2343                 if (ret)
2344                         return ret;
2345         }
2346
2347         if (new_hash) {
2348                 bkey_copy(&new_hash->key, new_key);
2349                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2350                                 new_hash, b->c.level, b->c.btree_id);
2351                 BUG_ON(ret);
2352         }
2353
2354         parent = btree_node_parent(btree_iter_path(trans, iter), b);
2355         if (parent) {
2356                 bch2_trans_copy_iter(&iter2, iter);
2357
2358                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2359                                 iter2.flags & BTREE_ITER_intent,
2360                                 _THIS_IP_);
2361
2362                 struct btree_path *path2 = btree_iter_path(trans, &iter2);
2363                 BUG_ON(path2->level != b->c.level);
2364                 BUG_ON(!bpos_eq(path2->pos, new_key->k.p));
2365
2366                 btree_path_set_level_up(trans, path2);
2367
2368                 trans->paths_sorted = false;
2369
2370                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2371                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_norun);
2372                 if (ret)
2373                         goto err;
2374         } else {
2375                 BUG_ON(btree_node_root(c, b) != b);
2376
2377                 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans,
2378                                        jset_u64s(new_key->k.u64s));
2379                 ret = PTR_ERR_OR_ZERO(e);
2380                 if (ret)
2381                         return ret;
2382
2383                 journal_entry_set(e,
2384                                   BCH_JSET_ENTRY_btree_root,
2385                                   b->c.btree_id, b->c.level,
2386                                   new_key, new_key->k.u64s);
2387         }
2388
2389         ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2390         if (ret)
2391                 goto err;
2392
2393         bch2_btree_node_lock_write_nofail(trans, btree_iter_path(trans, iter), &b->c);
2394
2395         if (new_hash) {
2396                 mutex_lock(&c->btree_cache.lock);
2397                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2398                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2399
2400                 bkey_copy(&b->key, new_key);
2401                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2402                 BUG_ON(ret);
2403                 mutex_unlock(&c->btree_cache.lock);
2404         } else {
2405                 bkey_copy(&b->key, new_key);
2406         }
2407
2408         bch2_btree_node_unlock_write(trans, btree_iter_path(trans, iter), b);
2409 out:
2410         bch2_trans_iter_exit(trans, &iter2);
2411         return ret;
2412 err:
2413         if (new_hash) {
2414                 mutex_lock(&c->btree_cache.lock);
2415                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2416                 mutex_unlock(&c->btree_cache.lock);
2417         }
2418         goto out;
2419 }
2420
2421 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2422                                struct btree *b, struct bkey_i *new_key,
2423                                unsigned commit_flags, bool skip_triggers)
2424 {
2425         struct bch_fs *c = trans->c;
2426         struct btree *new_hash = NULL;
2427         struct btree_path *path = btree_iter_path(trans, iter);
2428         struct closure cl;
2429         int ret = 0;
2430
2431         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2432         if (ret)
2433                 return ret;
2434
2435         closure_init_stack(&cl);
2436
2437         /*
2438          * check btree_ptr_hash_val() after @b is locked by
2439          * btree_iter_traverse():
2440          */
2441         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2442                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2443                 if (ret) {
2444                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2445                         if (ret)
2446                                 return ret;
2447                 }
2448
2449                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2450         }
2451
2452         path->intent_ref++;
2453         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2454                                            commit_flags, skip_triggers);
2455         --path->intent_ref;
2456
2457         if (new_hash) {
2458                 mutex_lock(&c->btree_cache.lock);
2459                 list_move(&new_hash->list, &c->btree_cache.freeable);
2460                 mutex_unlock(&c->btree_cache.lock);
2461
2462                 six_unlock_write(&new_hash->c.lock);
2463                 six_unlock_intent(&new_hash->c.lock);
2464         }
2465         closure_sync(&cl);
2466         bch2_btree_cache_cannibalize_unlock(trans);
2467         return ret;
2468 }
2469
2470 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2471                                         struct btree *b, struct bkey_i *new_key,
2472                                         unsigned commit_flags, bool skip_triggers)
2473 {
2474         struct btree_iter iter;
2475         int ret;
2476
2477         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2478                                   BTREE_MAX_DEPTH, b->c.level,
2479                                   BTREE_ITER_intent);
2480         ret = bch2_btree_iter_traverse(&iter);
2481         if (ret)
2482                 goto out;
2483
2484         /* has node been freed? */
2485         if (btree_iter_path(trans, &iter)->l[b->c.level].b != b) {
2486                 /* node has been freed: */
2487                 BUG_ON(!btree_node_dying(b));
2488                 goto out;
2489         }
2490
2491         BUG_ON(!btree_node_hashed(b));
2492
2493         bch2_bkey_drop_ptrs(bkey_i_to_s(new_key), ptr,
2494                             !bch2_bkey_has_device(bkey_i_to_s(&b->key), ptr->dev));
2495
2496         ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2497                                          commit_flags, skip_triggers);
2498 out:
2499         bch2_trans_iter_exit(trans, &iter);
2500         return ret;
2501 }
2502
2503 /* Init code: */
2504
2505 /*
2506  * Only for filesystem bringup, when first reading the btree roots or allocating
2507  * btree roots when initializing a new filesystem:
2508  */
2509 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2510 {
2511         BUG_ON(btree_node_root(c, b));
2512
2513         bch2_btree_set_root_inmem(c, b);
2514 }
2515
2516 int bch2_btree_root_alloc_fake_trans(struct btree_trans *trans, enum btree_id id, unsigned level)
2517 {
2518         struct bch_fs *c = trans->c;
2519         struct closure cl;
2520         struct btree *b;
2521         int ret;
2522
2523         closure_init_stack(&cl);
2524
2525         do {
2526                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2527                 closure_sync(&cl);
2528         } while (ret);
2529
2530         b = bch2_btree_node_mem_alloc(trans, false);
2531         bch2_btree_cache_cannibalize_unlock(trans);
2532
2533         set_btree_node_fake(b);
2534         set_btree_node_need_rewrite(b);
2535         b->c.level      = level;
2536         b->c.btree_id   = id;
2537
2538         bkey_btree_ptr_init(&b->key);
2539         b->key.k.p = SPOS_MAX;
2540         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2541
2542         bch2_bset_init_first(b, &b->data->keys);
2543         bch2_btree_build_aux_trees(b);
2544
2545         b->data->flags = 0;
2546         btree_set_min(b, POS_MIN);
2547         btree_set_max(b, SPOS_MAX);
2548         b->data->format = bch2_btree_calc_format(b);
2549         btree_node_set_format(b, b->data->format);
2550
2551         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2552                                           b->c.level, b->c.btree_id);
2553         BUG_ON(ret);
2554
2555         bch2_btree_set_root_inmem(c, b);
2556
2557         six_unlock_write(&b->c.lock);
2558         six_unlock_intent(&b->c.lock);
2559         return 0;
2560 }
2561
2562 void bch2_btree_root_alloc_fake(struct bch_fs *c, enum btree_id id, unsigned level)
2563 {
2564         bch2_trans_run(c, bch2_btree_root_alloc_fake_trans(trans, id, level));
2565 }
2566
2567 static void bch2_btree_update_to_text(struct printbuf *out, struct btree_update *as)
2568 {
2569         prt_printf(out, "%ps: ", (void *) as->ip_started);
2570         bch2_trans_commit_flags_to_text(out, as->flags);
2571
2572         prt_printf(out, " btree=%s l=%u-%u mode=%s nodes_written=%u cl.remaining=%u journal_seq=%llu\n",
2573                    bch2_btree_id_str(as->btree_id),
2574                    as->update_level_start,
2575                    as->update_level_end,
2576                    bch2_btree_update_modes[as->mode],
2577                    as->nodes_written,
2578                    closure_nr_remaining(&as->cl),
2579                    as->journal.seq);
2580 }
2581
2582 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2583 {
2584         struct btree_update *as;
2585
2586         mutex_lock(&c->btree_interior_update_lock);
2587         list_for_each_entry(as, &c->btree_interior_update_list, list)
2588                 bch2_btree_update_to_text(out, as);
2589         mutex_unlock(&c->btree_interior_update_lock);
2590 }
2591
2592 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2593 {
2594         bool ret;
2595
2596         mutex_lock(&c->btree_interior_update_lock);
2597         ret = !list_empty(&c->btree_interior_update_list);
2598         mutex_unlock(&c->btree_interior_update_lock);
2599
2600         return ret;
2601 }
2602
2603 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2604 {
2605         bool ret = bch2_btree_interior_updates_pending(c);
2606
2607         if (ret)
2608                 closure_wait_event(&c->btree_interior_update_wait,
2609                                    !bch2_btree_interior_updates_pending(c));
2610         return ret;
2611 }
2612
2613 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2614 {
2615         struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2616
2617         mutex_lock(&c->btree_root_lock);
2618
2619         r->level = entry->level;
2620         r->alive = true;
2621         bkey_copy(&r->key, (struct bkey_i *) entry->start);
2622
2623         mutex_unlock(&c->btree_root_lock);
2624 }
2625
2626 struct jset_entry *
2627 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2628                                     struct jset_entry *end,
2629                                     unsigned long skip)
2630 {
2631         unsigned i;
2632
2633         mutex_lock(&c->btree_root_lock);
2634
2635         for (i = 0; i < btree_id_nr_alive(c); i++) {
2636                 struct btree_root *r = bch2_btree_id_root(c, i);
2637
2638                 if (r->alive && !test_bit(i, &skip)) {
2639                         journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2640                                           i, r->level, &r->key, r->key.k.u64s);
2641                         end = vstruct_next(end);
2642                 }
2643         }
2644
2645         mutex_unlock(&c->btree_root_lock);
2646
2647         return end;
2648 }
2649
2650 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2651 {
2652         if (c->btree_node_rewrite_worker)
2653                 destroy_workqueue(c->btree_node_rewrite_worker);
2654         if (c->btree_interior_update_worker)
2655                 destroy_workqueue(c->btree_interior_update_worker);
2656         mempool_exit(&c->btree_interior_update_pool);
2657 }
2658
2659 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2660 {
2661         mutex_init(&c->btree_reserve_cache_lock);
2662         INIT_LIST_HEAD(&c->btree_interior_update_list);
2663         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2664         mutex_init(&c->btree_interior_update_lock);
2665         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2666
2667         INIT_LIST_HEAD(&c->pending_node_rewrites);
2668         mutex_init(&c->pending_node_rewrites_lock);
2669 }
2670
2671 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2672 {
2673         c->btree_interior_update_worker =
2674                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 8);
2675         if (!c->btree_interior_update_worker)
2676                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2677
2678         c->btree_node_rewrite_worker =
2679                 alloc_ordered_workqueue("btree_node_rewrite", WQ_UNBOUND);
2680         if (!c->btree_node_rewrite_worker)
2681                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2682
2683         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2684                                       sizeof(struct btree_update)))
2685                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2686
2687         return 0;
2688 }