git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_leaf.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_update_interior.h"
   6 #include "btree_gc.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_key_cache.h"
  10 #include "btree_locking.h"
  11 #include "buckets.h"
  12 #include "debug.h"
  13 #include "error.h"
  14 #include "extent_update.h"
  15 #include "journal.h"
  16 #include "journal_reclaim.h"
  17 #include "keylist.h"
  18 #include "replicas.h"
  19
  20 #include <linux/prefetch.h>
  21 #include <linux/sort.h>
  22 #include <trace/events/bcachefs.h>
  23
  24 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  25                                          const struct btree_insert_entry *r)
  26 {
  27         return   cmp_int(l->btree_id,   r->btree_id) ?:
  28                  -cmp_int(l->level,     r->level) ?:
  29                  bpos_cmp(l->k->k.p,    r->k->k.p);
  30 }
  31
  32 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  33                                      struct btree_insert_entry *i)
  34 {
  35         return i != trans->updates2 &&
  36                 iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
  37 }
  38
  39 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
  40                                             struct btree_iter *iter)
  41 {
  42         bch2_btree_node_lock_write(b, iter);
  43
  44         if (btree_iter_type(iter) == BTREE_ITER_CACHED)
  45                 return;
  46
  47         if (unlikely(btree_node_just_written(b)) &&
  48             bch2_btree_post_write_cleanup(c, b))
  49                 bch2_btree_iter_reinit_node(iter, b);
  50
  51         /*
  52          * If the last bset has been written, or if it's gotten too big - start
  53          * a new bset to insert into:
  54          */
  55         if (want_new_bset(c, b))
  56                 bch2_btree_init_next(c, b, iter);
  57 }
  58
  59 /* Inserting into a given leaf node (last stage of insert): */
  60
  61 /* Handle overwrites and do insert, for non extents: */
  62 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  63                                 struct btree *b,
  64                                 struct btree_node_iter *node_iter,
  65                                 struct bkey_i *insert)
  66 {
  67         struct bkey_packed *k;
  68         unsigned clobber_u64s = 0, new_u64s = 0;
  69
  70         EBUG_ON(btree_node_just_written(b));
  71         EBUG_ON(bset_written(b, btree_bset_last(b)));
  72         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  73         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
  74         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
  75         EBUG_ON(insert->k.u64s >
  76                 bch_btree_keys_u64s_remaining(iter->trans->c, b));
  77         EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
  78
  79         k = bch2_btree_node_iter_peek_all(node_iter, b);
  80         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
  81                 k = NULL;
  82
  83         /* @k is the key being overwritten/deleted, if any: */
  84         EBUG_ON(k && bkey_deleted(k));
  85
  86         /* Deleting, but not found? nothing to do: */
  87         if (bkey_deleted(&insert->k) && !k)
  88                 return false;
  89
  90         if (bkey_deleted(&insert->k)) {
  91                 /* Deleting: */
  92                 btree_account_key_drop(b, k);
  93                 k->type = KEY_TYPE_deleted;
  94
  95                 if (k->needs_whiteout)
  96                         push_whiteout(iter->trans->c, b, insert->k.p);
  97                 k->needs_whiteout = false;
  98
  99                 if (k >= btree_bset_last(b)->start) {
 100                         clobber_u64s = k->u64s;
 101                         bch2_bset_delete(b, k, clobber_u64s);
 102                         goto fix_iter;
 103                 } else {
 104                         bch2_btree_iter_fix_key_modified(iter, b, k);
 105                 }
 106
 107                 return true;
 108         }
 109
 110         if (k) {
 111                 /* Overwriting: */
 112                 btree_account_key_drop(b, k);
 113                 k->type = KEY_TYPE_deleted;
 114
 115                 insert->k.needs_whiteout = k->needs_whiteout;
 116                 k->needs_whiteout = false;
 117
 118                 if (k >= btree_bset_last(b)->start) {
 119                         clobber_u64s = k->u64s;
 120                         goto overwrite;
 121                 } else {
 122                         bch2_btree_iter_fix_key_modified(iter, b, k);
 123                 }
 124         }
 125
 126         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 127 overwrite:
 128         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 129         new_u64s = k->u64s;
 130 fix_iter:
 131         if (clobber_u64s != new_u64s)
 132                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 133                                          clobber_u64s, new_u64s);
 134         return true;
 135 }
 136
 137 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 138                                unsigned i, u64 seq)
 139 {
 140         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 141         struct btree_write *w = container_of(pin, struct btree_write, journal);
 142         struct btree *b = container_of(w, struct btree, writes[i]);
 143
 144         btree_node_lock_type(c, b, SIX_LOCK_read);
 145         bch2_btree_node_write_cond(c, b,
 146                 (btree_current_write(b) == w && w->journal.seq == seq));
 147         six_unlock_read(&b->c.lock);
 148         return 0;
 149 }
 150
 151 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 152 {
 153         return __btree_node_flush(j, pin, 0, seq);
 154 }
 155
 156 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 157 {
 158         return __btree_node_flush(j, pin, 1, seq);
 159 }
 160
 161 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 162                                        struct btree *b, u64 seq)
 163 {
 164         struct btree_write *w = btree_current_write(b);
 165
 166         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 167                              btree_node_write_idx(b) == 0
 168                              ? btree_node_flush0
 169                              : btree_node_flush1);
 170 }
 171
 172 /**
 173  * btree_insert_key - insert a key one key into a leaf node
 174  */
 175 static bool btree_insert_key_leaf(struct btree_trans *trans,
 176                                   struct btree_iter *iter,
 177                                   struct bkey_i *insert)
 178 {
 179         struct bch_fs *c = trans->c;
 180         struct btree *b = iter_l(iter)->b;
 181         struct bset_tree *t = bset_tree_last(b);
 182         struct bset *i = bset(b, t);
 183         int old_u64s = bset_u64s(t);
 184         int old_live_u64s = b->nr.live_u64s;
 185         int live_u64s_added, u64s_added;
 186
 187         EBUG_ON(!iter->level &&
 188                 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 189
 190         if (unlikely(!bch2_btree_bset_insert_key(iter, b,
 191                                         &iter_l(iter)->iter, insert)))
 192                 return false;
 193
 194         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 195                                          le64_to_cpu(i->journal_seq)));
 196
 197         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 198
 199         if (unlikely(!btree_node_dirty(b)))
 200                 set_btree_node_dirty(c, b);
 201
 202         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 203         u64s_added = (int) bset_u64s(t) - old_u64s;
 204
 205         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 206                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 207         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 208                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 209
 210         if (u64s_added > live_u64s_added &&
 211             bch2_maybe_compact_whiteouts(c, b))
 212                 bch2_btree_iter_reinit_node(iter, b);
 213
 214         trace_btree_insert_key(c, b, insert);
 215         return true;
 216 }
 217
 218 /* Cached btree updates: */
 219
 220 /* Normal update interface: */
 221
 222 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 223                                              struct btree_insert_entry *i)
 224 {
 225         BUG_ON(!i->is_extent && bpos_cmp(i->k->k.p, i->iter->real_pos));
 226         BUG_ON(i->level         != i->iter->level);
 227         BUG_ON(i->btree_id      != i->iter->btree_id);
 228 }
 229
 230 static noinline int
 231 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 232 {
 233         struct bch_fs *c = trans->c;
 234         int ret;
 235
 236         bch2_trans_unlock(trans);
 237
 238         ret = bch2_journal_preres_get(&c->journal,
 239                         &trans->journal_preres, u64s, 0);
 240         if (ret)
 241                 return ret;
 242
 243         if (!bch2_trans_relock(trans)) {
 244                 trace_trans_restart_journal_preres_get(trans->ip);
 245                 return -EINTR;
 246         }
 247
 248         return 0;
 249 }
 250
 251 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 252                                              unsigned flags)
 253 {
 254         struct bch_fs *c = trans->c;
 255         int ret;
 256
 257         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 258                 flags |= JOURNAL_RES_GET_RESERVED;
 259
 260         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 261                                    trans->journal_u64s, flags);
 262
 263         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 264 }
 265
 266 static enum btree_insert_ret
 267 btree_key_can_insert(struct btree_trans *trans,
 268                      struct btree_iter *iter,
 269                      unsigned u64s)
 270 {
 271         struct bch_fs *c = trans->c;
 272         struct btree *b = iter_l(iter)->b;
 273
 274         if (!bch2_btree_node_insert_fits(c, b, u64s))
 275                 return BTREE_INSERT_BTREE_NODE_FULL;
 276
 277         return BTREE_INSERT_OK;
 278 }
 279
 280 static enum btree_insert_ret
 281 btree_key_can_insert_cached(struct btree_trans *trans,
 282                             struct btree_iter *iter,
 283                             unsigned u64s)
 284 {
 285         struct bkey_cached *ck = (void *) iter->l[0].b;
 286         unsigned new_u64s;
 287         struct bkey_i *new_k;
 288
 289         BUG_ON(iter->level);
 290
 291         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 292             bch2_btree_key_cache_must_wait(trans->c) &&
 293             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 294                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 295
 296         /*
 297          * bch2_varint_decode can read past the end of the buffer by at most 7
 298          * bytes (it won't be used):
 299          */
 300         u64s += 1;
 301
 302         if (u64s <= ck->u64s)
 303                 return BTREE_INSERT_OK;
 304
 305         new_u64s        = roundup_pow_of_two(u64s);
 306         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 307         if (!new_k)
 308                 return -ENOMEM;
 309
 310         ck->u64s        = new_u64s;
 311         ck->k           = new_k;
 312         return BTREE_INSERT_OK;
 313 }
 314
 315 static inline void do_btree_insert_one(struct btree_trans *trans,
 316                                        struct btree_insert_entry *i)
 317 {
 318         struct bch_fs *c = trans->c;
 319         struct journal *j = &c->journal;
 320         bool did_work;
 321
 322         EBUG_ON(trans->journal_res.ref !=
 323                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 324
 325         i->k->k.needs_whiteout = false;
 326
 327         did_work = (btree_iter_type(i->iter) != BTREE_ITER_CACHED)
 328                 ? btree_insert_key_leaf(trans, i->iter, i->k)
 329                 : bch2_btree_insert_key_cached(trans, i->iter, i->k);
 330         if (!did_work)
 331                 return;
 332
 333         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 334                 bch2_journal_add_keys(j, &trans->journal_res,
 335                                       i->btree_id,
 336                                       i->level,
 337                                       i->k);
 338
 339                 bch2_journal_set_has_inode(j, &trans->journal_res,
 340                                            i->k->k.p.inode);
 341
 342                 if (trans->journal_seq)
 343                         *trans->journal_seq = trans->journal_res.seq;
 344         }
 345 }
 346
 347 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 348 {
 349         __bch2_btree_iter_unlock(iter);
 350 }
 351
 352 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 353 {
 354         struct bch_fs *c = trans->c;
 355         struct btree_insert_entry *i;
 356
 357         trans_for_each_update(trans, i) {
 358                 /*
 359                  * XXX: synchronization of cached update triggers with gc
 360                  */
 361                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 362
 363                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 364                         bch2_mark_update(trans, i->iter, i->k, NULL,
 365                                          i->trigger_flags|BTREE_TRIGGER_GC);
 366         }
 367 }
 368
 369 static inline int
 370 bch2_trans_commit_write_locked(struct btree_trans *trans,
 371                                struct btree_insert_entry **stopped_at)
 372 {
 373         struct bch_fs *c = trans->c;
 374         struct btree_insert_entry *i;
 375         struct btree_trans_commit_hook *h;
 376         unsigned u64s = 0;
 377         bool marking = false;
 378         int ret;
 379
 380         if (race_fault()) {
 381                 trace_trans_restart_fault_inject(trans->ip);
 382                 return -EINTR;
 383         }
 384
 385         /*
 386          * Check if the insert will fit in the leaf node with the write lock
 387          * held, otherwise another thread could write the node changing the
 388          * amount of space available:
 389          */
 390
 391         prefetch(&trans->c->journal.flags);
 392
 393         h = trans->hooks;
 394         while (h) {
 395                 ret = h->fn(trans, h);
 396                 if (ret)
 397                         return ret;
 398                 h = h->next;
 399         }
 400
 401         trans_for_each_update2(trans, i) {
 402                 /* Multiple inserts might go to same leaf: */
 403                 if (!same_leaf_as_prev(trans, i))
 404                         u64s = 0;
 405
 406                 u64s += i->k->k.u64s;
 407                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 408                         ? btree_key_can_insert(trans, i->iter, u64s)
 409                         : btree_key_can_insert_cached(trans, i->iter, u64s);
 410                 if (ret) {
 411                         *stopped_at = i;
 412                         return ret;
 413                 }
 414
 415                 if (btree_node_type_needs_gc(i->bkey_type))
 416                         marking = true;
 417         }
 418
 419         if (marking) {
 420                 percpu_down_read(&c->mark_lock);
 421         }
 422
 423         /* Must be called under mark_lock: */
 424         if (marking && trans->fs_usage_deltas &&
 425             !bch2_replicas_delta_list_marked(c, trans->fs_usage_deltas)) {
 426                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 427                 goto err;
 428         }
 429
 430         /*
 431          * Don't get journal reservation until after we know insert will
 432          * succeed:
 433          */
 434         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 435                 ret = bch2_trans_journal_res_get(trans,
 436                                 JOURNAL_RES_GET_NONBLOCK);
 437                 if (ret)
 438                         goto err;
 439         } else {
 440                 trans->journal_res.seq = c->journal.replay_journal_seq;
 441         }
 442
 443         if (unlikely(trans->extra_journal_entry_u64s)) {
 444                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 445                                   trans->extra_journal_entries,
 446                                   trans->extra_journal_entry_u64s);
 447
 448                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 449                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 450         }
 451
 452         /*
 453          * Not allowed to fail after we've gotten our journal reservation - we
 454          * have to use it:
 455          */
 456
 457         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 458                 if (bch2_journal_seq_verify)
 459                         trans_for_each_update2(trans, i)
 460                                 i->k->k.version.lo = trans->journal_res.seq;
 461                 else if (bch2_inject_invalid_keys)
 462                         trans_for_each_update2(trans, i)
 463                                 i->k->k.version = MAX_VERSION;
 464         }
 465
 466         trans_for_each_update(trans, i)
 467                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 468                         bch2_mark_update(trans, i->iter, i->k,
 469                                          NULL, i->trigger_flags);
 470
 471         if (marking && trans->fs_usage_deltas)
 472                 bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas);
 473
 474         if (unlikely(c->gc_pos.phase))
 475                 bch2_trans_mark_gc(trans);
 476
 477         trans_for_each_update2(trans, i)
 478                 do_btree_insert_one(trans, i);
 479 err:
 480         if (marking) {
 481                 percpu_up_read(&c->mark_lock);
 482         }
 483
 484         return ret;
 485 }
 486
 487 static noinline int maybe_do_btree_merge(struct btree_trans *trans, struct btree_iter *iter)
 488 {
 489         struct btree_insert_entry *i;
 490         struct btree *b = iter_l(iter)->b;
 491         struct bkey_s_c old;
 492         int u64s_delta = 0;
 493         int ret;
 494
 495         /*
 496          * Inserting directly into interior nodes is an uncommon operation with
 497          * various weird edge cases: also, a lot of things about
 498          * BTREE_ITER_NODES iters need to be audited
 499          */
 500         if (unlikely(btree_iter_type(iter) != BTREE_ITER_KEYS))
 501                 return 0;
 502
 503         BUG_ON(iter->level);
 504
 505         trans_for_each_update2(trans, i) {
 506                 if (iter_l(i->iter)->b != b)
 507                         continue;
 508
 509                 old = bch2_btree_iter_peek_slot(i->iter);
 510                 ret = bkey_err(old);
 511                 if (ret)
 512                         return ret;
 513
 514                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 515                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 516         }
 517
 518         return u64s_delta <= 0
 519                 ? (bch2_foreground_maybe_merge(trans->c, iter, iter->level,
 520                                 trans->flags & ~BTREE_INSERT_NOUNLOCK) ?: -EINTR)
 521                 : 0;
 522 }
 523
 524 /*
 525  * Get journal reservation, take write locks, and attempt to do btree update(s):
 526  */
 527 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 528                                        struct btree_insert_entry **stopped_at)
 529 {
 530         struct bch_fs *c = trans->c;
 531         struct btree_insert_entry *i;
 532         struct btree_iter *iter;
 533         int ret;
 534
 535         trans_for_each_update2(trans, i) {
 536                 struct btree *b;
 537
 538                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 539
 540                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED)
 541                         continue;
 542
 543                 b = iter_l(i->iter)->b;
 544                 if (b->sib_u64s[0] < c->btree_foreground_merge_threshold ||
 545                     b->sib_u64s[1] < c->btree_foreground_merge_threshold) {
 546                         ret = maybe_do_btree_merge(trans, i->iter);
 547                         if (unlikely(ret))
 548                                 return ret;
 549                 }
 550         }
 551
 552         trans_for_each_update2(trans, i)
 553                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 554
 555         ret = bch2_journal_preres_get(&c->journal,
 556                         &trans->journal_preres, trans->journal_preres_u64s,
 557                         JOURNAL_RES_GET_NONBLOCK|
 558                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 559                          ? JOURNAL_RES_GET_RESERVED : 0));
 560         if (unlikely(ret == -EAGAIN))
 561                 ret = bch2_trans_journal_preres_get_cold(trans,
 562                                                 trans->journal_preres_u64s);
 563         if (unlikely(ret))
 564                 return ret;
 565
 566         /*
 567          * Can't be holding any read locks when we go to take write locks:
 568          * another thread could be holding an intent lock on the same node we
 569          * have a read lock on, and it'll block trying to take a write lock
 570          * (because we hold a read lock) and it could be blocking us by holding
 571          * its own read lock (while we're trying to to take write locks).
 572          *
 573          * note - this must be done after bch2_trans_journal_preres_get_cold()
 574          * or anything else that might call bch2_trans_relock(), since that
 575          * would just retake the read locks:
 576          */
 577         trans_for_each_iter(trans, iter) {
 578                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 579                         if (btree_iter_keep(trans, iter)) {
 580                                 if (!bch2_btree_iter_upgrade(iter, 1)) {
 581                                         trace_trans_restart_upgrade(trans->ip);
 582                                         return -EINTR;
 583                                 }
 584                         } else {
 585                                 bch2_btree_iter_unlock_noinline(iter);
 586                         }
 587                 }
 588         }
 589
 590         trans_for_each_update2(trans, i) {
 591                 const char *invalid = bch2_bkey_invalid(c,
 592                                 bkey_i_to_s_c(i->k), i->bkey_type);
 593                 if (invalid) {
 594                         char buf[200];
 595
 596                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 597                         bch_err(c, "invalid bkey %s on insert: %s\n", buf, invalid);
 598                         bch2_fatal_error(c);
 599                 }
 600                 btree_insert_entry_checks(trans, i);
 601         }
 602         bch2_btree_trans_verify_locks(trans);
 603
 604         trans_for_each_update2(trans, i)
 605                 if (!same_leaf_as_prev(trans, i))
 606                         bch2_btree_node_lock_for_insert(c,
 607                                         iter_l(i->iter)->b, i->iter);
 608
 609         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 610
 611         trans_for_each_update2(trans, i)
 612                 if (!same_leaf_as_prev(trans, i))
 613                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 614                                                              i->iter);
 615
 616         if (!ret && trans->journal_pin)
 617                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 618                                      trans->journal_pin, NULL);
 619
 620         /*
 621          * Drop journal reservation after dropping write locks, since dropping
 622          * the journal reservation may kick off a journal write:
 623          */
 624         bch2_journal_res_put(&c->journal, &trans->journal_res);
 625
 626         if (unlikely(ret))
 627                 return ret;
 628
 629         bch2_trans_downgrade(trans);
 630
 631         return 0;
 632 }
 633
 634 static int journal_reclaim_wait_done(struct bch_fs *c)
 635 {
 636         int ret = bch2_journal_error(&c->journal) ?:
 637                 !bch2_btree_key_cache_must_wait(c);
 638
 639         if (!ret)
 640                 journal_reclaim_kick(&c->journal);
 641         return ret;
 642 }
 643
 644 static noinline
 645 int bch2_trans_commit_error(struct btree_trans *trans,
 646                             struct btree_insert_entry *i,
 647                             int ret)
 648 {
 649         struct bch_fs *c = trans->c;
 650         unsigned flags = trans->flags;
 651
 652         /*
 653          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 654          * update; if we haven't done anything yet it doesn't apply
 655          */
 656         flags &= ~BTREE_INSERT_NOUNLOCK;
 657
 658         switch (ret) {
 659         case BTREE_INSERT_BTREE_NODE_FULL:
 660                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 661
 662                 /*
 663                  * if the split succeeded without dropping locks the insert will
 664                  * still be atomic (what the caller peeked() and is overwriting
 665                  * won't have changed)
 666                  */
 667 #if 0
 668                 /*
 669                  * XXX:
 670                  * split -> btree node merging (of parent node) might still drop
 671                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 672                  *
 673                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 674                  * will inhibit merging - but we don't have a reliable way yet
 675                  * (do we?) of checking if we dropped locks in this path
 676                  */
 677                 if (!ret)
 678                         goto retry;
 679 #endif
 680
 681                 /*
 682                  * don't care if we got ENOSPC because we told split it
 683                  * couldn't block:
 684                  */
 685                 if (!ret ||
 686                     ret == -EINTR ||
 687                     (flags & BTREE_INSERT_NOUNLOCK)) {
 688                         trace_trans_restart_btree_node_split(trans->ip);
 689                         ret = -EINTR;
 690                 }
 691                 break;
 692         case BTREE_INSERT_ENOSPC:
 693                 ret = -ENOSPC;
 694                 break;
 695         case BTREE_INSERT_NEED_MARK_REPLICAS:
 696                 bch2_trans_unlock(trans);
 697
 698                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 699                 if (ret)
 700                         return ret;
 701
 702                 if (bch2_trans_relock(trans))
 703                         return 0;
 704
 705                 trace_trans_restart_mark_replicas(trans->ip);
 706                 ret = -EINTR;
 707                 break;
 708         case BTREE_INSERT_NEED_JOURNAL_RES:
 709                 bch2_trans_unlock(trans);
 710
 711                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 712                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED))
 713                         return -EAGAIN;
 714
 715                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 716                 if (ret)
 717                         return ret;
 718
 719                 if (bch2_trans_relock(trans))
 720                         return 0;
 721
 722                 trace_trans_restart_journal_res_get(trans->ip);
 723                 ret = -EINTR;
 724                 break;
 725         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 726                 bch2_trans_unlock(trans);
 727
 728                 wait_event_freezable(c->journal.reclaim_wait,
 729                                      (ret = journal_reclaim_wait_done(c)));
 730                 if (ret < 0)
 731                         return ret;
 732
 733                 if (bch2_trans_relock(trans))
 734                         return 0;
 735
 736                 trace_trans_restart_journal_reclaim(trans->ip);
 737                 ret = -EINTR;
 738                 break;
 739         default:
 740                 BUG_ON(ret >= 0);
 741                 break;
 742         }
 743
 744         return ret;
 745 }
 746
 747 static noinline int
 748 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 749 {
 750         struct bch_fs *c = trans->c;
 751         int ret;
 752
 753         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 754                 return -EROFS;
 755
 756         bch2_trans_unlock(trans);
 757
 758         ret = bch2_fs_read_write_early(c);
 759         if (ret)
 760                 return ret;
 761
 762         percpu_ref_get(&c->writes);
 763         return 0;
 764 }
 765
 766 static void __bch2_trans_update2(struct btree_trans *trans,
 767                                  struct btree_insert_entry n)
 768 {
 769         struct btree_insert_entry *i;
 770
 771         btree_insert_entry_checks(trans, &n);
 772
 773         EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX);
 774
 775         n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 776
 777         trans_for_each_update2(trans, i)
 778                 if (btree_insert_entry_cmp(&n, i) <= 0)
 779                         break;
 780
 781         if (i < trans->updates2 + trans->nr_updates2 &&
 782             !btree_insert_entry_cmp(&n, i))
 783                 *i = n;
 784         else
 785                 array_insert_item(trans->updates2, trans->nr_updates2,
 786                                   i - trans->updates2, n);
 787 }
 788
 789 static void bch2_trans_update2(struct btree_trans *trans,
 790                                struct btree_iter *iter,
 791                                struct bkey_i *insert)
 792 {
 793         __bch2_trans_update2(trans, (struct btree_insert_entry) {
 794                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
 795                 .btree_id       = iter->btree_id,
 796                 .level          = iter->level,
 797                 .iter           = iter,
 798                 .k              = insert,
 799         });
 800 }
 801
 802 static int extent_update_to_keys(struct btree_trans *trans,
 803                                  struct btree_insert_entry n)
 804 {
 805         int ret;
 806
 807         if (bkey_deleted(&n.k->k))
 808                 return 0;
 809
 810         ret = bch2_extent_can_insert(trans, n.iter, n.k);
 811         if (ret)
 812                 return ret;
 813
 814         n.iter = bch2_trans_get_iter(trans, n.iter->btree_id, n.k->k.p,
 815                                      BTREE_ITER_INTENT|
 816                                      BTREE_ITER_NOT_EXTENTS);
 817         n.is_extent = false;
 818
 819         __bch2_trans_update2(trans, n);
 820         bch2_trans_iter_put(trans, n.iter);
 821         return 0;
 822 }
 823
 824 static int extent_handle_overwrites(struct btree_trans *trans,
 825                                     enum btree_id btree_id,
 826                                     struct bkey_i *insert)
 827 {
 828         struct btree_iter *iter, *update_iter;
 829         struct bpos start = bkey_start_pos(&insert->k);
 830         struct bkey_i *update;
 831         struct bkey_s_c k;
 832         int ret = 0;
 833
 834         iter = bch2_trans_get_iter(trans, btree_id, start,
 835                                    BTREE_ITER_INTENT);
 836         k = bch2_btree_iter_peek_with_updates(iter);
 837
 838         while (k.k && !(ret = bkey_err(k))) {
 839                 if (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0)
 840                         break;
 841
 842                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 843                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 844                         if ((ret = PTR_ERR_OR_ZERO(update)))
 845                                 break;
 846
 847                         bkey_reassemble(update, k);
 848
 849                         bch2_cut_back(start, update);
 850
 851                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 852                                                           BTREE_ITER_NOT_EXTENTS|
 853                                                           BTREE_ITER_INTENT);
 854                         bch2_trans_update2(trans, update_iter, update);
 855                         bch2_trans_iter_put(trans, update_iter);
 856                 }
 857
 858                 if (bkey_cmp(k.k->p, insert->k.p) < 0 ||
 859                     (!bkey_cmp(k.k->p, insert->k.p) && bkey_deleted(&insert->k))) {
 860                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 861                         if ((ret = PTR_ERR_OR_ZERO(update)))
 862                                 break;
 863
 864                         bkey_init(&update->k);
 865                         update->k.p = k.k->p;
 866
 867                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 868                                                           BTREE_ITER_NOT_EXTENTS|
 869                                                           BTREE_ITER_INTENT);
 870                         bch2_trans_update2(trans, update_iter, update);
 871                         bch2_trans_iter_put(trans, update_iter);
 872                 }
 873
 874                 if (bkey_cmp(k.k->p, insert->k.p) > 0) {
 875                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 876                         if ((ret = PTR_ERR_OR_ZERO(update)))
 877                                 break;
 878
 879                         bkey_reassemble(update, k);
 880                         bch2_cut_front(insert->k.p, update);
 881
 882                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 883                                                           BTREE_ITER_NOT_EXTENTS|
 884                                                           BTREE_ITER_INTENT);
 885                         bch2_trans_update2(trans, update_iter, update);
 886                         bch2_trans_iter_put(trans, update_iter);
 887                         break;
 888                 }
 889
 890                 k = bch2_btree_iter_next_with_updates(iter);
 891         }
 892         bch2_trans_iter_put(trans, iter);
 893
 894         return ret;
 895 }
 896
 897 int __bch2_trans_commit(struct btree_trans *trans)
 898 {
 899         struct btree_insert_entry *i = NULL;
 900         struct btree_iter *iter;
 901         bool trans_trigger_run;
 902         unsigned u64s, reset_flags = 0;
 903         int ret = 0;
 904
 905         if (!trans->nr_updates)
 906                 goto out_reset;
 907
 908         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 909                 lockdep_assert_held(&trans->c->gc_lock);
 910
 911         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 912
 913         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 914         trans->journal_preres_u64s      = 0;
 915
 916         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 917             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 918                 ret = bch2_trans_commit_get_rw_cold(trans);
 919                 if (ret)
 920                         goto out_reset;
 921         }
 922
 923 #ifdef CONFIG_BCACHEFS_DEBUG
 924         trans_for_each_update(trans, i)
 925                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 926                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 927                         bch2_btree_key_cache_verify_clean(trans,
 928                                         i->btree_id, i->k->k.p);
 929 #endif
 930
 931         /*
 932          * Running triggers will append more updates to the list of updates as
 933          * we're walking it:
 934          */
 935         do {
 936                 trans_trigger_run = false;
 937
 938                 trans_for_each_update(trans, i) {
 939                         if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 940                             !i->trans_triggers_run) {
 941                                 i->trans_triggers_run = true;
 942                                 trans_trigger_run = true;
 943
 944                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 945                                                              i->trigger_flags);
 946                                 if (unlikely(ret)) {
 947                                         if (ret == -EINTR)
 948                                                 trace_trans_restart_mark(trans->ip);
 949                                         goto out;
 950                                 }
 951                         }
 952                 }
 953         } while (trans_trigger_run);
 954
 955         /* Turn extents updates into keys: */
 956         trans_for_each_update(trans, i)
 957                 if (i->is_extent) {
 958                         ret = extent_handle_overwrites(trans, i->btree_id, i->k);
 959                         if (unlikely(ret))
 960                                 goto out;
 961                 }
 962
 963         trans_for_each_update(trans, i) {
 964                 ret = i->is_extent
 965                         ? extent_update_to_keys(trans, *i)
 966                         : (__bch2_trans_update2(trans, *i), 0);
 967                 if (unlikely(ret))
 968                         goto out;
 969         }
 970
 971         trans_for_each_update2(trans, i) {
 972                 ret = bch2_btree_iter_traverse(i->iter);
 973                 if (unlikely(ret)) {
 974                         trace_trans_restart_traverse(trans->ip);
 975                         goto out;
 976                 }
 977
 978                 if (unlikely(!bch2_btree_iter_upgrade(i->iter, i->level + 1))) {
 979                         trace_trans_restart_upgrade(trans->ip);
 980                         ret = -EINTR;
 981                         goto out;
 982                 }
 983
 984                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 985
 986                 u64s = jset_u64s(i->k->k.u64s);
 987                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 988                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 989                         trans->journal_preres_u64s += u64s;
 990                 trans->journal_u64s += u64s;
 991         }
 992 retry:
 993         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 994
 995         ret = do_bch2_trans_commit(trans, &i);
 996
 997         /* make sure we didn't drop or screw up locks: */
 998         bch2_btree_trans_verify_locks(trans);
 999
1000         if (ret)
1001                 goto err;
1002
1003         trans_for_each_iter(trans, iter)
1004                 if (btree_iter_live(trans, iter) &&
1005                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT))
1006                         bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
1007 out:
1008         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
1009
1010         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1011                 percpu_ref_put(&trans->c->writes);
1012 out_reset:
1013         if (!ret)
1014                 reset_flags |= TRANS_RESET_NOTRAVERSE;
1015         if (!ret && (trans->flags & BTREE_INSERT_NOUNLOCK))
1016                 reset_flags |= TRANS_RESET_NOUNLOCK;
1017         bch2_trans_reset(trans, reset_flags);
1018
1019         return ret;
1020 err:
1021         ret = bch2_trans_commit_error(trans, i, ret);
1022         if (ret)
1023                 goto out;
1024
1025         goto retry;
1026 }
1027
1028 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1029                       struct bkey_i *k, enum btree_trigger_flags flags)
1030 {
1031         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
1032                 .trigger_flags  = flags,
1033                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
1034                 .btree_id       = iter->btree_id,
1035                 .level          = iter->level,
1036                 .is_extent      = (iter->flags & BTREE_ITER_IS_EXTENTS) != 0,
1037                 .iter           = iter,
1038                 .k              = k
1039         };
1040
1041         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1042
1043 #ifdef CONFIG_BCACHEFS_DEBUG
1044         BUG_ON(bkey_cmp(iter->pos,
1045                         n.is_extent ? bkey_start_pos(&k->k) : k->k.p));
1046
1047         trans_for_each_update(trans, i) {
1048                 BUG_ON(bkey_cmp(i->iter->pos,
1049                                 i->is_extent ? bkey_start_pos(&i->k->k) : i->k->k.p));
1050
1051                 BUG_ON(i != trans->updates &&
1052                        btree_insert_entry_cmp(i - 1, i) >= 0);
1053         }
1054 #endif
1055
1056         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1057
1058         if (n.is_extent) {
1059                 iter->pos_after_commit = k->k.p;
1060                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
1061         }
1062
1063         /*
1064          * Pending updates are kept sorted: first, find position of new update,
1065          * then delete/trim any updates the new update overwrites:
1066          */
1067         if (!n.is_extent) {
1068                 trans_for_each_update(trans, i)
1069                         if (btree_insert_entry_cmp(&n, i) <= 0)
1070                                 break;
1071
1072                 if (i < trans->updates + trans->nr_updates &&
1073                     !btree_insert_entry_cmp(&n, i))
1074                         *i = n;
1075                 else
1076                         array_insert_item(trans->updates, trans->nr_updates,
1077                                           i - trans->updates, n);
1078         } else {
1079                 trans_for_each_update(trans, i)
1080                         if (btree_insert_entry_cmp(&n, i) < 0)
1081                                 break;
1082
1083                 while (i > trans->updates &&
1084                        i[-1].btree_id == n.btree_id &&
1085                        bkey_cmp(bkey_start_pos(&n.k->k),
1086                                 bkey_start_pos(&i[-1].k->k)) <= 0) {
1087                         --i;
1088                         array_remove_item(trans->updates, trans->nr_updates,
1089                                           i - trans->updates);
1090                 }
1091
1092                 if (i > trans->updates &&
1093                     i[-1].btree_id == n.btree_id &&
1094                     bkey_cmp(bkey_start_pos(&n.k->k), i[-1].k->k.p) < 0)
1095                         bch2_cut_back(bkey_start_pos(&n.k->k), i[-1].k);
1096
1097                 if (i < trans->updates + trans->nr_updates &&
1098                     i->btree_id == n.btree_id &&
1099                     bkey_cmp(n.k->k.p, bkey_start_pos(&i->k->k)) > 0) {
1100                         /* We don't handle splitting extents here: */
1101                         BUG_ON(bkey_cmp(bkey_start_pos(&n.k->k),
1102                                         bkey_start_pos(&i->k->k)) > 0);
1103
1104                         /*
1105                          * When we have an extent that overwrites the start of another
1106                          * update, trimming that extent will mean the iterator's
1107                          * position has to change since the iterator position has to
1108                          * match the extent's start pos - but we don't want to change
1109                          * the iterator pos if some other code is using it, so we may
1110                          * need to clone it:
1111                          */
1112                         if (btree_iter_live(trans, i->iter)) {
1113                                 i->iter = bch2_trans_copy_iter(trans, i->iter);
1114
1115                                 i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1116                                 bch2_trans_iter_put(trans, i->iter);
1117                         }
1118
1119                         bch2_cut_front(n.k->k.p, i->k);
1120                         bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1121                 }
1122
1123                 array_insert_item(trans->updates, trans->nr_updates,
1124                                   i - trans->updates, n);
1125         }
1126
1127         return 0;
1128 }
1129
1130 void bch2_trans_commit_hook(struct btree_trans *trans,
1131                             struct btree_trans_commit_hook *h)
1132 {
1133         h->next = trans->hooks;
1134         trans->hooks = h;
1135 }
1136
1137 int __bch2_btree_insert(struct btree_trans *trans,
1138                         enum btree_id id, struct bkey_i *k)
1139 {
1140         struct btree_iter *iter;
1141         int ret;
1142
1143         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1144                                    BTREE_ITER_INTENT);
1145
1146         ret = bch2_trans_update(trans, iter, k, 0);
1147         bch2_trans_iter_put(trans, iter);
1148         return ret;
1149 }
1150
1151 /**
1152  * bch2_btree_insert - insert keys into the extent btree
1153  * @c:                  pointer to struct bch_fs
1154  * @id:                 btree to insert into
1155  * @insert_keys:        list of keys to insert
1156  * @hook:               insert callback
1157  */
1158 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1159                       struct bkey_i *k,
1160                       struct disk_reservation *disk_res,
1161                       u64 *journal_seq, int flags)
1162 {
1163         return bch2_trans_do(c, disk_res, journal_seq, flags,
1164                              __bch2_btree_insert(&trans, id, k));
1165 }
1166
1167 int bch2_btree_delete_at(struct btree_trans *trans,
1168                          struct btree_iter *iter, unsigned flags)
1169 {
1170         struct bkey_i k;
1171
1172         bkey_init(&k.k);
1173         k.k.p = iter->pos;
1174
1175         bch2_trans_update(trans, iter, &k, 0);
1176         return bch2_trans_commit(trans, NULL, NULL,
1177                                  BTREE_INSERT_NOFAIL|flags);
1178 }
1179
1180 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1181                                   struct bpos start, struct bpos end,
1182                                   u64 *journal_seq)
1183 {
1184         struct btree_iter *iter;
1185         struct bkey_s_c k;
1186         int ret = 0;
1187
1188         iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
1189 retry:
1190         while ((k = bch2_btree_iter_peek(iter)).k &&
1191                !(ret = bkey_err(k)) &&
1192                bkey_cmp(iter->pos, end) < 0) {
1193                 struct bkey_i delete;
1194
1195                 bch2_trans_begin(trans);
1196
1197                 bkey_init(&delete.k);
1198
1199                 /*
1200                  * This could probably be more efficient for extents:
1201                  */
1202
1203                 /*
1204                  * For extents, iter.pos won't necessarily be the same as
1205                  * bkey_start_pos(k.k) (for non extents they always will be the
1206                  * same). It's important that we delete starting from iter.pos
1207                  * because the range we want to delete could start in the middle
1208                  * of k.
1209                  *
1210                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1211                  * bkey_start_pos(k.k)).
1212                  */
1213                 delete.k.p = iter->pos;
1214
1215                 if (btree_node_type_is_extents(iter->btree_id)) {
1216                         unsigned max_sectors =
1217                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1218
1219                         /* create the biggest key we can */
1220                         bch2_key_resize(&delete.k, max_sectors);
1221                         bch2_cut_back(end, &delete);
1222
1223                         ret = bch2_extent_trim_atomic(&delete, iter);
1224                         if (ret)
1225                                 break;
1226                 }
1227
1228                 bch2_trans_update(trans, iter, &delete, 0);
1229                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1230                                         BTREE_INSERT_NOFAIL);
1231                 if (ret)
1232                         break;
1233
1234                 bch2_trans_cond_resched(trans);
1235         }
1236
1237         if (ret == -EINTR) {
1238                 ret = 0;
1239                 goto retry;
1240         }
1241
1242         bch2_trans_iter_free(trans, iter);
1243         return ret;
1244 }
1245
1246 /*
1247  * bch_btree_delete_range - delete everything within a given range
1248  *
1249  * Range is a half open interval - [start, end)
1250  */
1251 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1252                             struct bpos start, struct bpos end,
1253                             u64 *journal_seq)
1254 {
1255         return bch2_trans_do(c, NULL, journal_seq, 0,
1256                              bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
1257 }