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         if (u64s <= ck->u64s)
 297                 return BTREE_INSERT_OK;
 298
 299         new_u64s        = roundup_pow_of_two(u64s);
 300         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 301         if (!new_k)
 302                 return -ENOMEM;
 303
 304         ck->u64s        = new_u64s;
 305         ck->k           = new_k;
 306         return BTREE_INSERT_OK;
 307 }
 308
 309 static inline void do_btree_insert_one(struct btree_trans *trans,
 310                                        struct btree_insert_entry *i)
 311 {
 312         struct bch_fs *c = trans->c;
 313         struct journal *j = &c->journal;
 314         bool did_work;
 315
 316         EBUG_ON(trans->journal_res.ref !=
 317                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 318
 319         i->k->k.needs_whiteout = false;
 320
 321         did_work = (btree_iter_type(i->iter) != BTREE_ITER_CACHED)
 322                 ? btree_insert_key_leaf(trans, i->iter, i->k)
 323                 : bch2_btree_insert_key_cached(trans, i->iter, i->k);
 324         if (!did_work)
 325                 return;
 326
 327         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 328                 bch2_journal_add_keys(j, &trans->journal_res,
 329                                       i->btree_id,
 330                                       i->level,
 331                                       i->k);
 332
 333                 bch2_journal_set_has_inode(j, &trans->journal_res,
 334                                            i->k->k.p.inode);
 335
 336                 if (trans->journal_seq)
 337                         *trans->journal_seq = trans->journal_res.seq;
 338         }
 339 }
 340
 341 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 342 {
 343         __bch2_btree_iter_unlock(iter);
 344 }
 345
 346 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 347 {
 348         struct bch_fs *c = trans->c;
 349         struct btree_insert_entry *i;
 350
 351         trans_for_each_update(trans, i) {
 352                 /*
 353                  * XXX: synchronization of cached update triggers with gc
 354                  */
 355                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 356
 357                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 358                         bch2_mark_update(trans, i->iter, i->k, NULL,
 359                                          i->trigger_flags|BTREE_TRIGGER_GC);
 360         }
 361 }
 362
 363 static inline int
 364 bch2_trans_commit_write_locked(struct btree_trans *trans,
 365                                struct btree_insert_entry **stopped_at)
 366 {
 367         struct bch_fs *c = trans->c;
 368         struct btree_insert_entry *i;
 369         struct btree_trans_commit_hook *h;
 370         unsigned u64s = 0;
 371         bool marking = false;
 372         int ret;
 373
 374         if (race_fault()) {
 375                 trace_trans_restart_fault_inject(trans->ip);
 376                 return -EINTR;
 377         }
 378
 379         /*
 380          * Check if the insert will fit in the leaf node with the write lock
 381          * held, otherwise another thread could write the node changing the
 382          * amount of space available:
 383          */
 384
 385         prefetch(&trans->c->journal.flags);
 386
 387         h = trans->hooks;
 388         while (h) {
 389                 ret = h->fn(trans, h);
 390                 if (ret)
 391                         return ret;
 392                 h = h->next;
 393         }
 394
 395         trans_for_each_update2(trans, i) {
 396                 /* Multiple inserts might go to same leaf: */
 397                 if (!same_leaf_as_prev(trans, i))
 398                         u64s = 0;
 399
 400                 u64s += i->k->k.u64s;
 401                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 402                         ? btree_key_can_insert(trans, i->iter, u64s)
 403                         : btree_key_can_insert_cached(trans, i->iter, u64s);
 404                 if (ret) {
 405                         *stopped_at = i;
 406                         return ret;
 407                 }
 408
 409                 if (btree_node_type_needs_gc(i->bkey_type))
 410                         marking = true;
 411         }
 412
 413         if (marking) {
 414                 percpu_down_read(&c->mark_lock);
 415         }
 416
 417         /* Must be called under mark_lock: */
 418         if (marking && trans->fs_usage_deltas &&
 419             !bch2_replicas_delta_list_marked(c, trans->fs_usage_deltas)) {
 420                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 421                 goto err;
 422         }
 423
 424         /*
 425          * Don't get journal reservation until after we know insert will
 426          * succeed:
 427          */
 428         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 429                 ret = bch2_trans_journal_res_get(trans,
 430                                 JOURNAL_RES_GET_NONBLOCK);
 431                 if (ret)
 432                         goto err;
 433         } else {
 434                 trans->journal_res.seq = c->journal.replay_journal_seq;
 435         }
 436
 437         if (unlikely(trans->extra_journal_entry_u64s)) {
 438                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 439                                   trans->extra_journal_entries,
 440                                   trans->extra_journal_entry_u64s);
 441
 442                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 443                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 444         }
 445
 446         /*
 447          * Not allowed to fail after we've gotten our journal reservation - we
 448          * have to use it:
 449          */
 450
 451         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 452                 if (bch2_journal_seq_verify)
 453                         trans_for_each_update2(trans, i)
 454                                 i->k->k.version.lo = trans->journal_res.seq;
 455                 else if (bch2_inject_invalid_keys)
 456                         trans_for_each_update2(trans, i)
 457                                 i->k->k.version = MAX_VERSION;
 458         }
 459
 460         trans_for_each_update(trans, i)
 461                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 462                         bch2_mark_update(trans, i->iter, i->k,
 463                                          NULL, i->trigger_flags);
 464
 465         if (marking && trans->fs_usage_deltas)
 466                 bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas);
 467
 468         if (unlikely(c->gc_pos.phase))
 469                 bch2_trans_mark_gc(trans);
 470
 471         trans_for_each_update2(trans, i)
 472                 do_btree_insert_one(trans, i);
 473 err:
 474         if (marking) {
 475                 percpu_up_read(&c->mark_lock);
 476         }
 477
 478         return ret;
 479 }
 480
 481 static noinline int maybe_do_btree_merge(struct btree_trans *trans, struct btree_iter *iter)
 482 {
 483         struct btree_insert_entry *i;
 484         struct btree *b = iter_l(iter)->b;
 485         struct bkey_s_c old;
 486         int u64s_delta = 0;
 487         int ret;
 488
 489         /*
 490          * Inserting directly into interior nodes is an uncommon operation with
 491          * various weird edge cases: also, a lot of things about
 492          * BTREE_ITER_NODES iters need to be audited
 493          */
 494         if (unlikely(btree_iter_type(iter) != BTREE_ITER_KEYS))
 495                 return 0;
 496
 497         BUG_ON(iter->level);
 498
 499         trans_for_each_update2(trans, i) {
 500                 if (iter_l(i->iter)->b != b)
 501                         continue;
 502
 503                 old = bch2_btree_iter_peek_slot(i->iter);
 504                 ret = bkey_err(old);
 505                 if (ret)
 506                         return ret;
 507
 508                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 509                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 510         }
 511
 512         return u64s_delta <= 0
 513                 ? (bch2_foreground_maybe_merge(trans->c, iter, iter->level,
 514                                 trans->flags & ~BTREE_INSERT_NOUNLOCK) ?: -EINTR)
 515                 : 0;
 516 }
 517
 518 /*
 519  * Get journal reservation, take write locks, and attempt to do btree update(s):
 520  */
 521 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 522                                        struct btree_insert_entry **stopped_at)
 523 {
 524         struct bch_fs *c = trans->c;
 525         struct btree_insert_entry *i;
 526         struct btree_iter *iter;
 527         int ret;
 528
 529         trans_for_each_update2(trans, i) {
 530                 struct btree *b;
 531
 532                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 533
 534                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED)
 535                         continue;
 536
 537                 b = iter_l(i->iter)->b;
 538                 if (b->sib_u64s[0] < c->btree_foreground_merge_threshold ||
 539                     b->sib_u64s[1] < c->btree_foreground_merge_threshold) {
 540                         ret = maybe_do_btree_merge(trans, i->iter);
 541                         if (unlikely(ret))
 542                                 return ret;
 543                 }
 544         }
 545
 546         trans_for_each_update2(trans, i)
 547                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 548
 549         ret = bch2_journal_preres_get(&c->journal,
 550                         &trans->journal_preres, trans->journal_preres_u64s,
 551                         JOURNAL_RES_GET_NONBLOCK|
 552                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 553                          ? JOURNAL_RES_GET_RESERVED : 0));
 554         if (unlikely(ret == -EAGAIN))
 555                 ret = bch2_trans_journal_preres_get_cold(trans,
 556                                                 trans->journal_preres_u64s);
 557         if (unlikely(ret))
 558                 return ret;
 559
 560         /*
 561          * Can't be holding any read locks when we go to take write locks:
 562          * another thread could be holding an intent lock on the same node we
 563          * have a read lock on, and it'll block trying to take a write lock
 564          * (because we hold a read lock) and it could be blocking us by holding
 565          * its own read lock (while we're trying to to take write locks).
 566          *
 567          * note - this must be done after bch2_trans_journal_preres_get_cold()
 568          * or anything else that might call bch2_trans_relock(), since that
 569          * would just retake the read locks:
 570          */
 571         trans_for_each_iter(trans, iter) {
 572                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 573                         if (btree_iter_keep(trans, iter)) {
 574                                 if (!bch2_btree_iter_upgrade(iter, 1)) {
 575                                         trace_trans_restart_upgrade(trans->ip);
 576                                         return -EINTR;
 577                                 }
 578                         } else {
 579                                 bch2_btree_iter_unlock_noinline(iter);
 580                         }
 581                 }
 582         }
 583
 584         trans_for_each_update2(trans, i) {
 585                 const char *invalid = bch2_bkey_invalid(c,
 586                                 bkey_i_to_s_c(i->k), i->bkey_type);
 587                 if (invalid) {
 588                         char buf[200];
 589
 590                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 591                         bch_err(c, "invalid bkey %s on insert: %s\n", buf, invalid);
 592                         bch2_fatal_error(c);
 593                 }
 594                 btree_insert_entry_checks(trans, i);
 595         }
 596         bch2_btree_trans_verify_locks(trans);
 597
 598         trans_for_each_update2(trans, i)
 599                 if (!same_leaf_as_prev(trans, i))
 600                         bch2_btree_node_lock_for_insert(c,
 601                                         iter_l(i->iter)->b, i->iter);
 602
 603         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 604
 605         trans_for_each_update2(trans, i)
 606                 if (!same_leaf_as_prev(trans, i))
 607                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 608                                                              i->iter);
 609
 610         if (!ret && trans->journal_pin)
 611                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 612                                      trans->journal_pin, NULL);
 613
 614         /*
 615          * Drop journal reservation after dropping write locks, since dropping
 616          * the journal reservation may kick off a journal write:
 617          */
 618         bch2_journal_res_put(&c->journal, &trans->journal_res);
 619
 620         if (unlikely(ret))
 621                 return ret;
 622
 623         bch2_trans_downgrade(trans);
 624
 625         return 0;
 626 }
 627
 628 static int journal_reclaim_wait_done(struct bch_fs *c)
 629 {
 630         int ret = bch2_journal_error(&c->journal) ?:
 631                 !bch2_btree_key_cache_must_wait(c);
 632
 633         if (!ret)
 634                 journal_reclaim_kick(&c->journal);
 635         return ret;
 636 }
 637
 638 static noinline
 639 int bch2_trans_commit_error(struct btree_trans *trans,
 640                             struct btree_insert_entry *i,
 641                             int ret)
 642 {
 643         struct bch_fs *c = trans->c;
 644         unsigned flags = trans->flags;
 645
 646         /*
 647          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 648          * update; if we haven't done anything yet it doesn't apply
 649          */
 650         flags &= ~BTREE_INSERT_NOUNLOCK;
 651
 652         switch (ret) {
 653         case BTREE_INSERT_BTREE_NODE_FULL:
 654                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 655
 656                 /*
 657                  * if the split succeeded without dropping locks the insert will
 658                  * still be atomic (what the caller peeked() and is overwriting
 659                  * won't have changed)
 660                  */
 661 #if 0
 662                 /*
 663                  * XXX:
 664                  * split -> btree node merging (of parent node) might still drop
 665                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 666                  *
 667                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 668                  * will inhibit merging - but we don't have a reliable way yet
 669                  * (do we?) of checking if we dropped locks in this path
 670                  */
 671                 if (!ret)
 672                         goto retry;
 673 #endif
 674
 675                 /*
 676                  * don't care if we got ENOSPC because we told split it
 677                  * couldn't block:
 678                  */
 679                 if (!ret ||
 680                     ret == -EINTR ||
 681                     (flags & BTREE_INSERT_NOUNLOCK)) {
 682                         trace_trans_restart_btree_node_split(trans->ip);
 683                         ret = -EINTR;
 684                 }
 685                 break;
 686         case BTREE_INSERT_ENOSPC:
 687                 ret = -ENOSPC;
 688                 break;
 689         case BTREE_INSERT_NEED_MARK_REPLICAS:
 690                 bch2_trans_unlock(trans);
 691
 692                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 693                 if (ret)
 694                         return ret;
 695
 696                 if (bch2_trans_relock(trans))
 697                         return 0;
 698
 699                 trace_trans_restart_mark_replicas(trans->ip);
 700                 ret = -EINTR;
 701                 break;
 702         case BTREE_INSERT_NEED_JOURNAL_RES:
 703                 bch2_trans_unlock(trans);
 704
 705                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 706                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED))
 707                         return -EAGAIN;
 708
 709                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 710                 if (ret)
 711                         return ret;
 712
 713                 if (bch2_trans_relock(trans))
 714                         return 0;
 715
 716                 trace_trans_restart_journal_res_get(trans->ip);
 717                 ret = -EINTR;
 718                 break;
 719         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 720                 bch2_trans_unlock(trans);
 721
 722                 wait_event_freezable(c->journal.reclaim_wait,
 723                                      (ret = journal_reclaim_wait_done(c)));
 724                 if (ret < 0)
 725                         return ret;
 726
 727                 if (bch2_trans_relock(trans))
 728                         return 0;
 729
 730                 trace_trans_restart_journal_reclaim(trans->ip);
 731                 ret = -EINTR;
 732                 break;
 733         default:
 734                 BUG_ON(ret >= 0);
 735                 break;
 736         }
 737
 738         return ret;
 739 }
 740
 741 static noinline int
 742 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 743 {
 744         struct bch_fs *c = trans->c;
 745         int ret;
 746
 747         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 748                 return -EROFS;
 749
 750         bch2_trans_unlock(trans);
 751
 752         ret = bch2_fs_read_write_early(c);
 753         if (ret)
 754                 return ret;
 755
 756         percpu_ref_get(&c->writes);
 757         return 0;
 758 }
 759
 760 static void __bch2_trans_update2(struct btree_trans *trans,
 761                                  struct btree_insert_entry n)
 762 {
 763         struct btree_insert_entry *i;
 764
 765         btree_insert_entry_checks(trans, &n);
 766
 767         EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX);
 768
 769         n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 770
 771         trans_for_each_update2(trans, i)
 772                 if (btree_insert_entry_cmp(&n, i) <= 0)
 773                         break;
 774
 775         if (i < trans->updates2 + trans->nr_updates2 &&
 776             !btree_insert_entry_cmp(&n, i))
 777                 *i = n;
 778         else
 779                 array_insert_item(trans->updates2, trans->nr_updates2,
 780                                   i - trans->updates2, n);
 781 }
 782
 783 static void bch2_trans_update2(struct btree_trans *trans,
 784                                struct btree_iter *iter,
 785                                struct bkey_i *insert)
 786 {
 787         __bch2_trans_update2(trans, (struct btree_insert_entry) {
 788                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
 789                 .btree_id       = iter->btree_id,
 790                 .level          = iter->level,
 791                 .iter           = iter,
 792                 .k              = insert,
 793         });
 794 }
 795
 796 static int extent_update_to_keys(struct btree_trans *trans,
 797                                  struct btree_insert_entry n)
 798 {
 799         int ret;
 800
 801         if (bkey_deleted(&n.k->k))
 802                 return 0;
 803
 804         ret = bch2_extent_can_insert(trans, n.iter, n.k);
 805         if (ret)
 806                 return ret;
 807
 808         n.iter = bch2_trans_get_iter(trans, n.iter->btree_id, n.k->k.p,
 809                                      BTREE_ITER_INTENT|
 810                                      BTREE_ITER_NOT_EXTENTS);
 811         n.is_extent = false;
 812
 813         __bch2_trans_update2(trans, n);
 814         bch2_trans_iter_put(trans, n.iter);
 815         return 0;
 816 }
 817
 818 static int extent_handle_overwrites(struct btree_trans *trans,
 819                                     enum btree_id btree_id,
 820                                     struct bkey_i *insert)
 821 {
 822         struct btree_iter *iter, *update_iter;
 823         struct bpos start = bkey_start_pos(&insert->k);
 824         struct bkey_i *update;
 825         struct bkey_s_c k;
 826         int ret = 0;
 827
 828         iter = bch2_trans_get_iter(trans, btree_id, start,
 829                                    BTREE_ITER_INTENT);
 830         k = bch2_btree_iter_peek_with_updates(iter);
 831
 832         while (k.k && !(ret = bkey_err(k))) {
 833                 if (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0)
 834                         break;
 835
 836                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 837                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 838                         if ((ret = PTR_ERR_OR_ZERO(update)))
 839                                 break;
 840
 841                         bkey_reassemble(update, k);
 842
 843                         bch2_cut_back(start, update);
 844
 845                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 846                                                           BTREE_ITER_NOT_EXTENTS|
 847                                                           BTREE_ITER_INTENT);
 848                         bch2_trans_update2(trans, update_iter, update);
 849                         bch2_trans_iter_put(trans, update_iter);
 850                 }
 851
 852                 if (bkey_cmp(k.k->p, insert->k.p) < 0 ||
 853                     (!bkey_cmp(k.k->p, insert->k.p) && bkey_deleted(&insert->k))) {
 854                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 855                         if ((ret = PTR_ERR_OR_ZERO(update)))
 856                                 break;
 857
 858                         bkey_init(&update->k);
 859                         update->k.p = k.k->p;
 860
 861                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 862                                                           BTREE_ITER_NOT_EXTENTS|
 863                                                           BTREE_ITER_INTENT);
 864                         bch2_trans_update2(trans, update_iter, update);
 865                         bch2_trans_iter_put(trans, update_iter);
 866                 }
 867
 868                 if (bkey_cmp(k.k->p, insert->k.p) > 0) {
 869                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 870                         if ((ret = PTR_ERR_OR_ZERO(update)))
 871                                 break;
 872
 873                         bkey_reassemble(update, k);
 874                         bch2_cut_front(insert->k.p, update);
 875
 876                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 877                                                           BTREE_ITER_NOT_EXTENTS|
 878                                                           BTREE_ITER_INTENT);
 879                         bch2_trans_update2(trans, update_iter, update);
 880                         bch2_trans_iter_put(trans, update_iter);
 881                         break;
 882                 }
 883
 884                 k = bch2_btree_iter_next_with_updates(iter);
 885         }
 886         bch2_trans_iter_put(trans, iter);
 887
 888         return ret;
 889 }
 890
 891 int __bch2_trans_commit(struct btree_trans *trans)
 892 {
 893         struct btree_insert_entry *i = NULL;
 894         struct btree_iter *iter;
 895         bool trans_trigger_run;
 896         unsigned u64s, reset_flags = 0;
 897         int ret = 0;
 898
 899         if (!trans->nr_updates)
 900                 goto out_reset;
 901
 902         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 903                 lockdep_assert_held(&trans->c->gc_lock);
 904
 905         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 906
 907         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 908         trans->journal_preres_u64s      = 0;
 909
 910         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 911             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 912                 ret = bch2_trans_commit_get_rw_cold(trans);
 913                 if (ret)
 914                         goto out_reset;
 915         }
 916
 917 #ifdef CONFIG_BCACHEFS_DEBUG
 918         trans_for_each_update(trans, i)
 919                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 920                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 921                         bch2_btree_key_cache_verify_clean(trans,
 922                                         i->btree_id, i->k->k.p);
 923 #endif
 924
 925         /*
 926          * Running triggers will append more updates to the list of updates as
 927          * we're walking it:
 928          */
 929         do {
 930                 trans_trigger_run = false;
 931
 932                 trans_for_each_update(trans, i) {
 933                         if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 934                             !i->trans_triggers_run) {
 935                                 i->trans_triggers_run = true;
 936                                 trans_trigger_run = true;
 937
 938                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 939                                                              i->trigger_flags);
 940                                 if (unlikely(ret)) {
 941                                         if (ret == -EINTR)
 942                                                 trace_trans_restart_mark(trans->ip);
 943                                         goto out;
 944                                 }
 945                         }
 946                 }
 947         } while (trans_trigger_run);
 948
 949         /* Turn extents updates into keys: */
 950         trans_for_each_update(trans, i)
 951                 if (i->is_extent) {
 952                         ret = extent_handle_overwrites(trans, i->btree_id, i->k);
 953                         if (unlikely(ret))
 954                                 goto out;
 955                 }
 956
 957         trans_for_each_update(trans, i) {
 958                 ret = i->is_extent
 959                         ? extent_update_to_keys(trans, *i)
 960                         : (__bch2_trans_update2(trans, *i), 0);
 961                 if (unlikely(ret))
 962                         goto out;
 963         }
 964
 965         trans_for_each_update2(trans, i) {
 966                 ret = bch2_btree_iter_traverse(i->iter);
 967                 if (unlikely(ret)) {
 968                         trace_trans_restart_traverse(trans->ip);
 969                         goto out;
 970                 }
 971
 972                 if (unlikely(!bch2_btree_iter_upgrade(i->iter, i->level + 1))) {
 973                         trace_trans_restart_upgrade(trans->ip);
 974                         ret = -EINTR;
 975                         goto out;
 976                 }
 977
 978                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 979
 980                 u64s = jset_u64s(i->k->k.u64s);
 981                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 982                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 983                         trans->journal_preres_u64s += u64s;
 984                 trans->journal_u64s += u64s;
 985         }
 986 retry:
 987         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 988
 989         ret = do_bch2_trans_commit(trans, &i);
 990
 991         /* make sure we didn't drop or screw up locks: */
 992         bch2_btree_trans_verify_locks(trans);
 993
 994         if (ret)
 995                 goto err;
 996
 997         trans_for_each_iter(trans, iter)
 998                 if (btree_iter_live(trans, iter) &&
 999                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT))
1000                         bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
1001 out:
1002         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
1003
1004         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1005                 percpu_ref_put(&trans->c->writes);
1006 out_reset:
1007         if (!ret)
1008                 reset_flags |= TRANS_RESET_NOTRAVERSE;
1009         if (!ret && (trans->flags & BTREE_INSERT_NOUNLOCK))
1010                 reset_flags |= TRANS_RESET_NOUNLOCK;
1011         bch2_trans_reset(trans, reset_flags);
1012
1013         return ret;
1014 err:
1015         ret = bch2_trans_commit_error(trans, i, ret);
1016         if (ret)
1017                 goto out;
1018
1019         goto retry;
1020 }
1021
1022 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1023                       struct bkey_i *k, enum btree_trigger_flags flags)
1024 {
1025         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
1026                 .trigger_flags  = flags,
1027                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
1028                 .btree_id       = iter->btree_id,
1029                 .level          = iter->level,
1030                 .is_extent      = (iter->flags & BTREE_ITER_IS_EXTENTS) != 0,
1031                 .iter           = iter,
1032                 .k              = k
1033         };
1034
1035         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1036
1037 #ifdef CONFIG_BCACHEFS_DEBUG
1038         BUG_ON(bkey_cmp(iter->pos,
1039                         n.is_extent ? bkey_start_pos(&k->k) : k->k.p));
1040
1041         trans_for_each_update(trans, i) {
1042                 BUG_ON(bkey_cmp(i->iter->pos,
1043                                 i->is_extent ? bkey_start_pos(&i->k->k) : i->k->k.p));
1044
1045                 BUG_ON(i != trans->updates &&
1046                        btree_insert_entry_cmp(i - 1, i) >= 0);
1047         }
1048 #endif
1049
1050         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1051
1052         if (n.is_extent) {
1053                 iter->pos_after_commit = k->k.p;
1054                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
1055         }
1056
1057         /*
1058          * Pending updates are kept sorted: first, find position of new update,
1059          * then delete/trim any updates the new update overwrites:
1060          */
1061         if (!n.is_extent) {
1062                 trans_for_each_update(trans, i)
1063                         if (btree_insert_entry_cmp(&n, i) <= 0)
1064                                 break;
1065
1066                 if (i < trans->updates + trans->nr_updates &&
1067                     !btree_insert_entry_cmp(&n, i))
1068                         *i = n;
1069                 else
1070                         array_insert_item(trans->updates, trans->nr_updates,
1071                                           i - trans->updates, n);
1072         } else {
1073                 trans_for_each_update(trans, i)
1074                         if (btree_insert_entry_cmp(&n, i) < 0)
1075                                 break;
1076
1077                 while (i > trans->updates &&
1078                        i[-1].btree_id == n.btree_id &&
1079                        bkey_cmp(bkey_start_pos(&n.k->k),
1080                                 bkey_start_pos(&i[-1].k->k)) <= 0) {
1081                         --i;
1082                         array_remove_item(trans->updates, trans->nr_updates,
1083                                           i - trans->updates);
1084                 }
1085
1086                 if (i > trans->updates &&
1087                     i[-1].btree_id == n.btree_id &&
1088                     bkey_cmp(bkey_start_pos(&n.k->k), i[-1].k->k.p) < 0)
1089                         bch2_cut_back(bkey_start_pos(&n.k->k), i[-1].k);
1090
1091                 if (i < trans->updates + trans->nr_updates &&
1092                     i->btree_id == n.btree_id &&
1093                     bkey_cmp(n.k->k.p, bkey_start_pos(&i->k->k)) > 0) {
1094                         /* We don't handle splitting extents here: */
1095                         BUG_ON(bkey_cmp(bkey_start_pos(&n.k->k),
1096                                         bkey_start_pos(&i->k->k)) > 0);
1097
1098                         /*
1099                          * When we have an extent that overwrites the start of another
1100                          * update, trimming that extent will mean the iterator's
1101                          * position has to change since the iterator position has to
1102                          * match the extent's start pos - but we don't want to change
1103                          * the iterator pos if some other code is using it, so we may
1104                          * need to clone it:
1105                          */
1106                         if (btree_iter_live(trans, i->iter)) {
1107                                 i->iter = bch2_trans_copy_iter(trans, i->iter);
1108
1109                                 i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1110                                 bch2_trans_iter_put(trans, i->iter);
1111                         }
1112
1113                         bch2_cut_front(n.k->k.p, i->k);
1114                         bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1115                 }
1116
1117                 array_insert_item(trans->updates, trans->nr_updates,
1118                                   i - trans->updates, n);
1119         }
1120
1121         return 0;
1122 }
1123
1124 void bch2_trans_commit_hook(struct btree_trans *trans,
1125                             struct btree_trans_commit_hook *h)
1126 {
1127         h->next = trans->hooks;
1128         trans->hooks = h;
1129 }
1130
1131 int __bch2_btree_insert(struct btree_trans *trans,
1132                         enum btree_id id, struct bkey_i *k)
1133 {
1134         struct btree_iter *iter;
1135         int ret;
1136
1137         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1138                                    BTREE_ITER_INTENT);
1139
1140         ret = bch2_trans_update(trans, iter, k, 0);
1141         bch2_trans_iter_put(trans, iter);
1142         return ret;
1143 }
1144
1145 /**
1146  * bch2_btree_insert - insert keys into the extent btree
1147  * @c:                  pointer to struct bch_fs
1148  * @id:                 btree to insert into
1149  * @insert_keys:        list of keys to insert
1150  * @hook:               insert callback
1151  */
1152 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1153                       struct bkey_i *k,
1154                       struct disk_reservation *disk_res,
1155                       u64 *journal_seq, int flags)
1156 {
1157         return bch2_trans_do(c, disk_res, journal_seq, flags,
1158                              __bch2_btree_insert(&trans, id, k));
1159 }
1160
1161 int bch2_btree_delete_at(struct btree_trans *trans,
1162                          struct btree_iter *iter, unsigned flags)
1163 {
1164         struct bkey_i k;
1165
1166         bkey_init(&k.k);
1167         k.k.p = iter->pos;
1168
1169         bch2_trans_update(trans, iter, &k, 0);
1170         return bch2_trans_commit(trans, NULL, NULL,
1171                                  BTREE_INSERT_NOFAIL|flags);
1172 }
1173
1174 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1175                                   struct bpos start, struct bpos end,
1176                                   u64 *journal_seq)
1177 {
1178         struct btree_iter *iter;
1179         struct bkey_s_c k;
1180         int ret = 0;
1181
1182         iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
1183 retry:
1184         while ((k = bch2_btree_iter_peek(iter)).k &&
1185                !(ret = bkey_err(k)) &&
1186                bkey_cmp(iter->pos, end) < 0) {
1187                 struct bkey_i delete;
1188
1189                 bch2_trans_begin(trans);
1190
1191                 bkey_init(&delete.k);
1192
1193                 /*
1194                  * This could probably be more efficient for extents:
1195                  */
1196
1197                 /*
1198                  * For extents, iter.pos won't necessarily be the same as
1199                  * bkey_start_pos(k.k) (for non extents they always will be the
1200                  * same). It's important that we delete starting from iter.pos
1201                  * because the range we want to delete could start in the middle
1202                  * of k.
1203                  *
1204                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1205                  * bkey_start_pos(k.k)).
1206                  */
1207                 delete.k.p = iter->pos;
1208
1209                 if (btree_node_type_is_extents(iter->btree_id)) {
1210                         unsigned max_sectors =
1211                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1212
1213                         /* create the biggest key we can */
1214                         bch2_key_resize(&delete.k, max_sectors);
1215                         bch2_cut_back(end, &delete);
1216
1217                         ret = bch2_extent_trim_atomic(&delete, iter);
1218                         if (ret)
1219                                 break;
1220                 }
1221
1222                 bch2_trans_update(trans, iter, &delete, 0);
1223                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1224                                         BTREE_INSERT_NOFAIL);
1225                 if (ret)
1226                         break;
1227
1228                 bch2_trans_cond_resched(trans);
1229         }
1230
1231         if (ret == -EINTR) {
1232                 ret = 0;
1233                 goto retry;
1234         }
1235
1236         bch2_trans_iter_free(trans, iter);
1237         return ret;
1238 }
1239
1240 /*
1241  * bch_btree_delete_range - delete everything within a given range
1242  *
1243  * Range is a half open interval - [start, end)
1244  */
1245 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1246                             struct bpos start, struct bpos end,
1247                             u64 *journal_seq)
1248 {
1249         return bch2_trans_do(c, NULL, journal_seq, 0,
1250                              bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
1251 }