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->updates &&
  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 btree_trans *trans,
  40                                             struct btree_iter *iter,
  41                                             struct btree *b)
  42 {
  43         struct bch_fs *c = trans->c;
  44
  45         bch2_btree_node_lock_write(b, iter);
  46
  47         if (btree_iter_type(iter) == BTREE_ITER_CACHED)
  48                 return;
  49
  50         if (unlikely(btree_node_just_written(b)) &&
  51             bch2_btree_post_write_cleanup(c, b))
  52                 bch2_btree_iter_reinit_node(iter, b);
  53
  54         /*
  55          * If the last bset has been written, or if it's gotten too big - start
  56          * a new bset to insert into:
  57          */
  58         if (want_new_bset(c, b))
  59                 bch2_btree_init_next(trans, iter, b);
  60 }
  61
  62 /* Inserting into a given leaf node (last stage of insert): */
  63
  64 /* Handle overwrites and do insert, for non extents: */
  65 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  66                                 struct btree *b,
  67                                 struct btree_node_iter *node_iter,
  68                                 struct bkey_i *insert)
  69 {
  70         struct bkey_packed *k;
  71         unsigned clobber_u64s = 0, new_u64s = 0;
  72
  73         EBUG_ON(btree_node_just_written(b));
  74         EBUG_ON(bset_written(b, btree_bset_last(b)));
  75         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  76         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
  77         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
  78         EBUG_ON(insert->k.u64s >
  79                 bch_btree_keys_u64s_remaining(iter->trans->c, b));
  80         EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
  81
  82         k = bch2_btree_node_iter_peek_all(node_iter, b);
  83         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
  84                 k = NULL;
  85
  86         /* @k is the key being overwritten/deleted, if any: */
  87         EBUG_ON(k && bkey_deleted(k));
  88
  89         /* Deleting, but not found? nothing to do: */
  90         if (bkey_deleted(&insert->k) && !k)
  91                 return false;
  92
  93         if (bkey_deleted(&insert->k)) {
  94                 /* Deleting: */
  95                 btree_account_key_drop(b, k);
  96                 k->type = KEY_TYPE_deleted;
  97
  98                 if (k->needs_whiteout)
  99                         push_whiteout(iter->trans->c, b, insert->k.p);
 100                 k->needs_whiteout = false;
 101
 102                 if (k >= btree_bset_last(b)->start) {
 103                         clobber_u64s = k->u64s;
 104                         bch2_bset_delete(b, k, clobber_u64s);
 105                         goto fix_iter;
 106                 } else {
 107                         bch2_btree_iter_fix_key_modified(iter, b, k);
 108                 }
 109
 110                 return true;
 111         }
 112
 113         if (k) {
 114                 /* Overwriting: */
 115                 btree_account_key_drop(b, k);
 116                 k->type = KEY_TYPE_deleted;
 117
 118                 insert->k.needs_whiteout = k->needs_whiteout;
 119                 k->needs_whiteout = false;
 120
 121                 if (k >= btree_bset_last(b)->start) {
 122                         clobber_u64s = k->u64s;
 123                         goto overwrite;
 124                 } else {
 125                         bch2_btree_iter_fix_key_modified(iter, b, k);
 126                 }
 127         }
 128
 129         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 130 overwrite:
 131         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 132         new_u64s = k->u64s;
 133 fix_iter:
 134         if (clobber_u64s != new_u64s)
 135                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 136                                          clobber_u64s, new_u64s);
 137         return true;
 138 }
 139
 140 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 141                                unsigned i, u64 seq)
 142 {
 143         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 144         struct btree_write *w = container_of(pin, struct btree_write, journal);
 145         struct btree *b = container_of(w, struct btree, writes[i]);
 146
 147         btree_node_lock_type(c, b, SIX_LOCK_read);
 148         bch2_btree_node_write_cond(c, b,
 149                 (btree_current_write(b) == w && w->journal.seq == seq));
 150         six_unlock_read(&b->c.lock);
 151         return 0;
 152 }
 153
 154 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 155 {
 156         return __btree_node_flush(j, pin, 0, seq);
 157 }
 158
 159 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 160 {
 161         return __btree_node_flush(j, pin, 1, seq);
 162 }
 163
 164 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 165                                        struct btree *b, u64 seq)
 166 {
 167         struct btree_write *w = btree_current_write(b);
 168
 169         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 170                              btree_node_write_idx(b) == 0
 171                              ? btree_node_flush0
 172                              : btree_node_flush1);
 173 }
 174
 175 /**
 176  * btree_insert_key - insert a key one key into a leaf node
 177  */
 178 static bool btree_insert_key_leaf(struct btree_trans *trans,
 179                                   struct btree_iter *iter,
 180                                   struct bkey_i *insert)
 181 {
 182         struct bch_fs *c = trans->c;
 183         struct btree *b = iter_l(iter)->b;
 184         struct bset_tree *t = bset_tree_last(b);
 185         struct bset *i = bset(b, t);
 186         int old_u64s = bset_u64s(t);
 187         int old_live_u64s = b->nr.live_u64s;
 188         int live_u64s_added, u64s_added;
 189
 190         EBUG_ON(!iter->level &&
 191                 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 192
 193         if (unlikely(!bch2_btree_bset_insert_key(iter, b,
 194                                         &iter_l(iter)->iter, insert)))
 195                 return false;
 196
 197         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 198                                          le64_to_cpu(i->journal_seq)));
 199
 200         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 201
 202         if (unlikely(!btree_node_dirty(b)))
 203                 set_btree_node_dirty(c, b);
 204
 205         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 206         u64s_added = (int) bset_u64s(t) - old_u64s;
 207
 208         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 209                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 210         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 211                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 212
 213         if (u64s_added > live_u64s_added &&
 214             bch2_maybe_compact_whiteouts(c, b))
 215                 bch2_btree_iter_reinit_node(iter, b);
 216
 217         trace_btree_insert_key(c, b, insert);
 218         return true;
 219 }
 220
 221 /* Cached btree updates: */
 222
 223 /* Normal update interface: */
 224
 225 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 226                                              struct btree_insert_entry *i)
 227 {
 228         BUG_ON(bpos_cmp(i->k->k.p, i->iter->real_pos));
 229         BUG_ON(i->level         != i->iter->level);
 230         BUG_ON(i->btree_id      != i->iter->btree_id);
 231 }
 232
 233 static noinline int
 234 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 235                                    unsigned long trace_ip)
 236 {
 237         struct bch_fs *c = trans->c;
 238         int ret;
 239
 240         bch2_trans_unlock(trans);
 241
 242         ret = bch2_journal_preres_get(&c->journal,
 243                         &trans->journal_preres, u64s, 0);
 244         if (ret)
 245                 return ret;
 246
 247         if (!bch2_trans_relock(trans)) {
 248                 trace_trans_restart_journal_preres_get(trans->ip, trace_ip);
 249                 return -EINTR;
 250         }
 251
 252         return 0;
 253 }
 254
 255 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 256                                              unsigned flags)
 257 {
 258         struct bch_fs *c = trans->c;
 259         int ret;
 260
 261         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 262                 flags |= JOURNAL_RES_GET_RESERVED;
 263
 264         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 265                                    trans->journal_u64s, flags);
 266
 267         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 268 }
 269
 270 static enum btree_insert_ret
 271 btree_key_can_insert(struct btree_trans *trans,
 272                      struct btree_iter *iter,
 273                      unsigned u64s)
 274 {
 275         struct bch_fs *c = trans->c;
 276         struct btree *b = iter_l(iter)->b;
 277
 278         if (!bch2_btree_node_insert_fits(c, b, u64s))
 279                 return BTREE_INSERT_BTREE_NODE_FULL;
 280
 281         return BTREE_INSERT_OK;
 282 }
 283
 284 static enum btree_insert_ret
 285 btree_key_can_insert_cached(struct btree_trans *trans,
 286                             struct btree_iter *iter,
 287                             unsigned u64s)
 288 {
 289         struct bkey_cached *ck = (void *) iter->l[0].b;
 290         unsigned new_u64s;
 291         struct bkey_i *new_k;
 292
 293         BUG_ON(iter->level);
 294
 295         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 296             bch2_btree_key_cache_must_wait(trans->c) &&
 297             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 298                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 299
 300         /*
 301          * bch2_varint_decode can read past the end of the buffer by at most 7
 302          * bytes (it won't be used):
 303          */
 304         u64s += 1;
 305
 306         if (u64s <= ck->u64s)
 307                 return BTREE_INSERT_OK;
 308
 309         new_u64s        = roundup_pow_of_two(u64s);
 310         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 311         if (!new_k)
 312                 return -ENOMEM;
 313
 314         ck->u64s        = new_u64s;
 315         ck->k           = new_k;
 316         return BTREE_INSERT_OK;
 317 }
 318
 319 static inline void do_btree_insert_one(struct btree_trans *trans,
 320                                        struct btree_insert_entry *i)
 321 {
 322         struct bch_fs *c = trans->c;
 323         struct journal *j = &c->journal;
 324         bool did_work;
 325
 326         EBUG_ON(trans->journal_res.ref !=
 327                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 328
 329         i->k->k.needs_whiteout = false;
 330
 331         did_work = (btree_iter_type(i->iter) != BTREE_ITER_CACHED)
 332                 ? btree_insert_key_leaf(trans, i->iter, i->k)
 333                 : bch2_btree_insert_key_cached(trans, i->iter, i->k);
 334         if (!did_work)
 335                 return;
 336
 337         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 338                 bch2_journal_add_keys(j, &trans->journal_res,
 339                                       i->btree_id,
 340                                       i->level,
 341                                       i->k);
 342
 343                 bch2_journal_set_has_inode(j, &trans->journal_res,
 344                                            i->k->k.p.inode);
 345
 346                 if (trans->journal_seq)
 347                         *trans->journal_seq = trans->journal_res.seq;
 348         }
 349 }
 350
 351 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 352 {
 353         __bch2_btree_iter_unlock(iter);
 354 }
 355
 356 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 357 {
 358         struct bch_fs *c = trans->c;
 359         struct btree_insert_entry *i;
 360
 361         trans_for_each_update(trans, i) {
 362                 /*
 363                  * XXX: synchronization of cached update triggers with gc
 364                  */
 365                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 366
 367                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 368                         bch2_mark_update(trans, i->iter, i->k,
 369                                          i->flags|BTREE_TRIGGER_GC);
 370         }
 371 }
 372
 373 static inline int
 374 bch2_trans_commit_write_locked(struct btree_trans *trans,
 375                                struct btree_insert_entry **stopped_at,
 376                                unsigned long trace_ip)
 377 {
 378         struct bch_fs *c = trans->c;
 379         struct btree_insert_entry *i;
 380         struct btree_trans_commit_hook *h;
 381         unsigned u64s = 0;
 382         bool marking = false;
 383         int ret;
 384
 385         if (race_fault()) {
 386                 trace_trans_restart_fault_inject(trans->ip, trace_ip);
 387                 return -EINTR;
 388         }
 389
 390         /*
 391          * Check if the insert will fit in the leaf node with the write lock
 392          * held, otherwise another thread could write the node changing the
 393          * amount of space available:
 394          */
 395
 396         prefetch(&trans->c->journal.flags);
 397
 398         h = trans->hooks;
 399         while (h) {
 400                 ret = h->fn(trans, h);
 401                 if (ret)
 402                         return ret;
 403                 h = h->next;
 404         }
 405
 406         trans_for_each_update(trans, i) {
 407                 /* Multiple inserts might go to same leaf: */
 408                 if (!same_leaf_as_prev(trans, i))
 409                         u64s = 0;
 410
 411                 u64s += i->k->k.u64s;
 412                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 413                         ? btree_key_can_insert(trans, i->iter, u64s)
 414                         : btree_key_can_insert_cached(trans, i->iter, u64s);
 415                 if (ret) {
 416                         *stopped_at = i;
 417                         return ret;
 418                 }
 419
 420                 if (btree_node_type_needs_gc(i->bkey_type))
 421                         marking = true;
 422         }
 423
 424         if (marking) {
 425                 percpu_down_read(&c->mark_lock);
 426         }
 427
 428         /* Must be called under mark_lock: */
 429         if (marking && trans->fs_usage_deltas &&
 430             !bch2_replicas_delta_list_marked(c, trans->fs_usage_deltas)) {
 431                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 432                 goto err;
 433         }
 434
 435         /*
 436          * Don't get journal reservation until after we know insert will
 437          * succeed:
 438          */
 439         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 440                 ret = bch2_trans_journal_res_get(trans,
 441                                 JOURNAL_RES_GET_NONBLOCK);
 442                 if (ret)
 443                         goto err;
 444         } else {
 445                 trans->journal_res.seq = c->journal.replay_journal_seq;
 446         }
 447
 448         if (unlikely(trans->extra_journal_entry_u64s)) {
 449                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 450                                   trans->extra_journal_entries,
 451                                   trans->extra_journal_entry_u64s);
 452
 453                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 454                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 455         }
 456
 457         /*
 458          * Not allowed to fail after we've gotten our journal reservation - we
 459          * have to use it:
 460          */
 461
 462         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 463                 if (bch2_journal_seq_verify)
 464                         trans_for_each_update(trans, i)
 465                                 i->k->k.version.lo = trans->journal_res.seq;
 466                 else if (bch2_inject_invalid_keys)
 467                         trans_for_each_update(trans, i)
 468                                 i->k->k.version = MAX_VERSION;
 469         }
 470
 471         trans_for_each_update(trans, i)
 472                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 473                         bch2_mark_update(trans, i->iter, i->k,
 474                                          i->flags);
 475
 476         if (marking && trans->fs_usage_deltas)
 477                 bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas);
 478
 479         if (unlikely(c->gc_pos.phase))
 480                 bch2_trans_mark_gc(trans);
 481
 482         trans_for_each_update(trans, i)
 483                 do_btree_insert_one(trans, i);
 484 err:
 485         if (marking) {
 486                 percpu_up_read(&c->mark_lock);
 487         }
 488
 489         return ret;
 490 }
 491
 492 static noinline int maybe_do_btree_merge(struct btree_trans *trans, struct btree_iter *iter)
 493 {
 494         struct btree_insert_entry *i;
 495         struct btree *b = iter_l(iter)->b;
 496         struct bkey_s_c old;
 497         int u64s_delta = 0;
 498         int ret;
 499
 500         /*
 501          * Inserting directly into interior nodes is an uncommon operation with
 502          * various weird edge cases: also, a lot of things about
 503          * BTREE_ITER_NODES iters need to be audited
 504          */
 505         if (unlikely(btree_iter_type(iter) != BTREE_ITER_KEYS))
 506                 return 0;
 507
 508         BUG_ON(iter->level);
 509
 510         trans_for_each_update(trans, i) {
 511                 if (iter_l(i->iter)->b != b)
 512                         continue;
 513
 514                 old = bch2_btree_iter_peek_slot(i->iter);
 515                 ret = bkey_err(old);
 516                 if (ret)
 517                         return ret;
 518
 519                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 520                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 521         }
 522
 523         return u64s_delta <= 0
 524                 ? (bch2_foreground_maybe_merge(trans, iter, iter->level,
 525                                 trans->flags & ~BTREE_INSERT_NOUNLOCK) ?: -EINTR)
 526                 : 0;
 527 }
 528
 529 /*
 530  * Get journal reservation, take write locks, and attempt to do btree update(s):
 531  */
 532 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 533                                        struct btree_insert_entry **stopped_at,
 534                                        unsigned long trace_ip)
 535 {
 536         struct bch_fs *c = trans->c;
 537         struct btree_insert_entry *i;
 538         struct btree_iter *iter;
 539         int ret;
 540
 541         trans_for_each_update(trans, i) {
 542                 struct btree *b;
 543
 544                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 545
 546                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED)
 547                         continue;
 548
 549                 b = iter_l(i->iter)->b;
 550                 if (b->sib_u64s[0] < c->btree_foreground_merge_threshold ||
 551                     b->sib_u64s[1] < c->btree_foreground_merge_threshold) {
 552                         ret = maybe_do_btree_merge(trans, i->iter);
 553                         if (unlikely(ret))
 554                                 return ret;
 555                 }
 556         }
 557
 558         trans_for_each_update(trans, i)
 559                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 560
 561         ret = bch2_journal_preres_get(&c->journal,
 562                         &trans->journal_preres, trans->journal_preres_u64s,
 563                         JOURNAL_RES_GET_NONBLOCK|
 564                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 565                          ? JOURNAL_RES_GET_RESERVED : 0));
 566         if (unlikely(ret == -EAGAIN))
 567                 ret = bch2_trans_journal_preres_get_cold(trans,
 568                                                 trans->journal_preres_u64s, trace_ip);
 569         if (unlikely(ret))
 570                 return ret;
 571
 572         /*
 573          * Can't be holding any read locks when we go to take write locks:
 574          * another thread could be holding an intent lock on the same node we
 575          * have a read lock on, and it'll block trying to take a write lock
 576          * (because we hold a read lock) and it could be blocking us by holding
 577          * its own read lock (while we're trying to to take write locks).
 578          *
 579          * note - this must be done after bch2_trans_journal_preres_get_cold()
 580          * or anything else that might call bch2_trans_relock(), since that
 581          * would just retake the read locks:
 582          */
 583         trans_for_each_iter(trans, iter) {
 584                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 585                         if (btree_iter_keep(trans, iter)) {
 586                                 if (!bch2_btree_iter_upgrade(iter, 1)) {
 587                                         trace_trans_restart_upgrade(trans->ip, trace_ip,
 588                                                                     iter->btree_id,
 589                                                                     &iter->real_pos);
 590                                         return -EINTR;
 591                                 }
 592                         } else {
 593                                 bch2_btree_iter_unlock_noinline(iter);
 594                         }
 595                 }
 596         }
 597
 598         trans_for_each_update(trans, i) {
 599                 const char *invalid = bch2_bkey_invalid(c,
 600                                 bkey_i_to_s_c(i->k), i->bkey_type);
 601                 if (invalid) {
 602                         char buf[200];
 603
 604                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 605                         bch_err(c, "invalid bkey %s on insert: %s\n", buf, invalid);
 606                         bch2_fatal_error(c);
 607                 }
 608                 btree_insert_entry_checks(trans, i);
 609         }
 610         bch2_btree_trans_verify_locks(trans);
 611
 612         trans_for_each_update(trans, i)
 613                 if (!same_leaf_as_prev(trans, i))
 614                         bch2_btree_node_lock_for_insert(trans, i->iter,
 615                                         iter_l(i->iter)->b);
 616
 617         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 618
 619         trans_for_each_update(trans, i)
 620                 if (!same_leaf_as_prev(trans, i))
 621                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 622                                                              i->iter);
 623
 624         if (!ret && trans->journal_pin)
 625                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 626                                      trans->journal_pin, NULL);
 627
 628         /*
 629          * Drop journal reservation after dropping write locks, since dropping
 630          * the journal reservation may kick off a journal write:
 631          */
 632         bch2_journal_res_put(&c->journal, &trans->journal_res);
 633
 634         if (unlikely(ret))
 635                 return ret;
 636
 637         bch2_trans_downgrade(trans);
 638
 639         return 0;
 640 }
 641
 642 static int journal_reclaim_wait_done(struct bch_fs *c)
 643 {
 644         int ret = bch2_journal_error(&c->journal) ?:
 645                 !bch2_btree_key_cache_must_wait(c);
 646
 647         if (!ret)
 648                 journal_reclaim_kick(&c->journal);
 649         return ret;
 650 }
 651
 652 static noinline
 653 int bch2_trans_commit_error(struct btree_trans *trans,
 654                             struct btree_insert_entry *i,
 655                             int ret, unsigned long trace_ip)
 656 {
 657         struct bch_fs *c = trans->c;
 658         unsigned flags = trans->flags;
 659
 660         /*
 661          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 662          * update; if we haven't done anything yet it doesn't apply
 663          */
 664         flags &= ~BTREE_INSERT_NOUNLOCK;
 665
 666         switch (ret) {
 667         case BTREE_INSERT_BTREE_NODE_FULL:
 668                 ret = bch2_btree_split_leaf(trans, i->iter, flags);
 669
 670                 /*
 671                  * if the split succeeded without dropping locks the insert will
 672                  * still be atomic (what the caller peeked() and is overwriting
 673                  * won't have changed)
 674                  */
 675 #if 0
 676                 /*
 677                  * XXX:
 678                  * split -> btree node merging (of parent node) might still drop
 679                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 680                  *
 681                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 682                  * will inhibit merging - but we don't have a reliable way yet
 683                  * (do we?) of checking if we dropped locks in this path
 684                  */
 685                 if (!ret)
 686                         goto retry;
 687 #endif
 688
 689                 /*
 690                  * don't care if we got ENOSPC because we told split it
 691                  * couldn't block:
 692                  */
 693                 if (!ret ||
 694                     ret == -EINTR ||
 695                     (flags & BTREE_INSERT_NOUNLOCK)) {
 696                         trace_trans_restart_btree_node_split(trans->ip, trace_ip,
 697                                                              i->iter->btree_id,
 698                                                              &i->iter->real_pos);
 699                         ret = -EINTR;
 700                 }
 701                 break;
 702         case BTREE_INSERT_NEED_MARK_REPLICAS:
 703                 bch2_trans_unlock(trans);
 704
 705                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 706                 if (ret)
 707                         return ret;
 708
 709                 if (bch2_trans_relock(trans))
 710                         return 0;
 711
 712                 trace_trans_restart_mark_replicas(trans->ip, trace_ip);
 713                 ret = -EINTR;
 714                 break;
 715         case BTREE_INSERT_NEED_JOURNAL_RES:
 716                 bch2_trans_unlock(trans);
 717
 718                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 719                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED))
 720                         return -EAGAIN;
 721
 722                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 723                 if (ret)
 724                         return ret;
 725
 726                 if (bch2_trans_relock(trans))
 727                         return 0;
 728
 729                 trace_trans_restart_journal_res_get(trans->ip, trace_ip);
 730                 ret = -EINTR;
 731                 break;
 732         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 733                 bch2_trans_unlock(trans);
 734
 735                 trace_trans_blocked_journal_reclaim(trans->ip, trace_ip);
 736
 737                 wait_event_freezable(c->journal.reclaim_wait,
 738                                      (ret = journal_reclaim_wait_done(c)));
 739                 if (ret < 0)
 740                         return ret;
 741
 742                 if (bch2_trans_relock(trans))
 743                         return 0;
 744
 745                 trace_trans_restart_journal_reclaim(trans->ip, trace_ip);
 746                 ret = -EINTR;
 747                 break;
 748         default:
 749                 BUG_ON(ret >= 0);
 750                 break;
 751         }
 752
 753         BUG_ON(ret == -ENOSPC && (flags & BTREE_INSERT_NOFAIL));
 754
 755         return ret;
 756 }
 757
 758 static noinline int
 759 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 760 {
 761         struct bch_fs *c = trans->c;
 762         int ret;
 763
 764         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 765                 return -EROFS;
 766
 767         bch2_trans_unlock(trans);
 768
 769         ret = bch2_fs_read_write_early(c);
 770         if (ret)
 771                 return ret;
 772
 773         percpu_ref_get(&c->writes);
 774         return 0;
 775 }
 776
 777 static int extent_handle_overwrites(struct btree_trans *trans,
 778                                     struct btree_insert_entry *i)
 779 {
 780         struct bch_fs *c = trans->c;
 781         struct btree_iter *iter, *update_iter;
 782         struct bpos start = bkey_start_pos(&i->k->k);
 783         struct bkey_i *update;
 784         struct bkey_s_c k;
 785         int ret = 0, compressed_sectors;
 786
 787         iter = bch2_trans_get_iter(trans, i->btree_id, start,
 788                                    BTREE_ITER_INTENT|
 789                                    BTREE_ITER_WITH_UPDATES|
 790                                    BTREE_ITER_NOT_EXTENTS);
 791         k = bch2_btree_iter_peek(iter);
 792         if (!k.k || (ret = bkey_err(k)))
 793                 goto out;
 794
 795         if (bch2_bkey_maybe_mergable(k.k, &i->k->k)) {
 796                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 797                 if ((ret = PTR_ERR_OR_ZERO(update)))
 798                         goto out;
 799
 800                 bkey_reassemble(update, k);
 801
 802                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(i->k))) {
 803                         update_iter = bch2_trans_copy_iter(trans, iter);
 804                         ret = bch2_btree_delete_at(trans, update_iter, i->flags);
 805                         bch2_trans_iter_put(trans, update_iter);
 806
 807                         if (ret)
 808                                 goto out;
 809
 810                         i->k = update;
 811                         goto next;
 812                 }
 813         }
 814
 815         if (!bkey_cmp(k.k->p, bkey_start_pos(&i->k->k)))
 816                 goto next;
 817
 818         while (bkey_cmp(i->k->k.p, bkey_start_pos(k.k)) > 0) {
 819                 /*
 820                  * If we're going to be splitting a compressed extent, note it
 821                  * so that __bch2_trans_commit() can increase our disk
 822                  * reservation:
 823                  */
 824                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0 &&
 825                     bkey_cmp(k.k->p, i->k->k.p) > 0 &&
 826                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
 827                         trans->extra_journal_res += compressed_sectors;
 828
 829                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 830                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 831                         if ((ret = PTR_ERR_OR_ZERO(update)))
 832                                 goto out;
 833
 834                         bkey_reassemble(update, k);
 835
 836                         bch2_cut_back(start, update);
 837
 838                         update_iter = bch2_trans_get_iter(trans, i->btree_id, update->k.p,
 839                                                           BTREE_ITER_NOT_EXTENTS|
 840                                                           BTREE_ITER_INTENT);
 841                         ret = bch2_btree_iter_traverse(update_iter);
 842                         if (ret)
 843                                 goto out;
 844
 845                         bch2_trans_update(trans, update_iter, update,
 846                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 847                                           i->flags);
 848                         bch2_trans_iter_put(trans, update_iter);
 849                 }
 850
 851                 if (bkey_cmp(k.k->p, i->k->k.p) <= 0) {
 852                         update_iter = bch2_trans_copy_iter(trans, iter);
 853                         ret = bch2_btree_delete_at(trans, update_iter,
 854                                                    i->flags);
 855                         bch2_trans_iter_put(trans, update_iter);
 856
 857                         if (ret)
 858                                 goto out;
 859                 }
 860
 861                 if (bkey_cmp(k.k->p, i->k->k.p) > 0) {
 862                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 863                         if ((ret = PTR_ERR_OR_ZERO(update)))
 864                                 goto out;
 865
 866                         bkey_reassemble(update, k);
 867                         bch2_cut_front(i->k->k.p, update);
 868
 869                         bch2_trans_update(trans, iter, update, i->flags);
 870                         goto out;
 871                 }
 872 next:
 873                 k = bch2_btree_iter_next(iter);
 874                 if (!k.k || (ret = bkey_err(k)))
 875                         goto out;
 876         }
 877
 878         bch2_bkey_merge(c, bkey_i_to_s(i->k), k);
 879 out:
 880         bch2_trans_iter_put(trans, iter);
 881
 882         return ret;
 883 }
 884
 885 int __bch2_trans_commit(struct btree_trans *trans)
 886 {
 887         struct btree_insert_entry *i = NULL;
 888         struct btree_iter *iter;
 889         bool trans_trigger_run;
 890         unsigned u64s, reset_flags = 0;
 891         int ret = 0;
 892
 893         if (!trans->nr_updates &&
 894             !trans->extra_journal_entry_u64s)
 895                 goto out_reset;
 896
 897         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 898                 lockdep_assert_held(&trans->c->gc_lock);
 899
 900         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 901
 902         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 903         trans->journal_preres_u64s      = 0;
 904
 905         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 906             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 907                 ret = bch2_trans_commit_get_rw_cold(trans);
 908                 if (ret)
 909                         goto out_reset;
 910         }
 911
 912 #ifdef CONFIG_BCACHEFS_DEBUG
 913         trans_for_each_update(trans, i)
 914                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 915                     !(i->flags & BTREE_TRIGGER_NORUN))
 916                         bch2_btree_key_cache_verify_clean(trans,
 917                                         i->btree_id, i->k->k.p);
 918 #endif
 919
 920         /*
 921          * Running triggers will append more updates to the list of updates as
 922          * we're walking it:
 923          */
 924         do {
 925                 trans_trigger_run = false;
 926
 927                 trans_for_each_update(trans, i) {
 928                         if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 929                             !i->trans_triggers_run) {
 930                                 i->trans_triggers_run = true;
 931                                 trans_trigger_run = true;
 932
 933                                 ret = bch2_trans_mark_update(trans, i->iter,
 934                                                              i->k, i->flags);
 935                                 if (unlikely(ret)) {
 936                                         if (ret == -EINTR)
 937                                                 trace_trans_restart_mark(trans->ip, _RET_IP_,
 938                                                                          i->iter->btree_id,
 939                                                                          &i->iter->pos);
 940                                         goto out;
 941                                 }
 942                         }
 943                 }
 944         } while (trans_trigger_run);
 945
 946         trans_for_each_update(trans, i) {
 947                 ret = bch2_btree_iter_traverse(i->iter);
 948                 if (unlikely(ret)) {
 949                         trace_trans_restart_traverse(trans->ip, _RET_IP_,
 950                                                      i->iter->btree_id,
 951                                                      &i->iter->pos);
 952                         goto out;
 953                 }
 954
 955                 if (unlikely(!bch2_btree_iter_upgrade(i->iter, i->level + 1))) {
 956                         trace_trans_restart_upgrade(trans->ip, _RET_IP_,
 957                                                     i->iter->btree_id,
 958                                                     &i->iter->pos);
 959                         ret = -EINTR;
 960                         goto out;
 961                 }
 962
 963                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 964
 965                 u64s = jset_u64s(i->k->k.u64s);
 966                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 967                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 968                         trans->journal_preres_u64s += u64s;
 969                 trans->journal_u64s += u64s;
 970         }
 971
 972         if (trans->extra_journal_res) {
 973                 ret = bch2_disk_reservation_add(trans->c, trans->disk_res,
 974                                 trans->extra_journal_res,
 975                                 (trans->flags & BTREE_INSERT_NOFAIL)
 976                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 977                 if (ret)
 978                         goto err;
 979         }
 980 retry:
 981         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 982
 983         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
 984
 985         /* make sure we didn't drop or screw up locks: */
 986         bch2_btree_trans_verify_locks(trans);
 987
 988         if (ret)
 989                 goto err;
 990
 991         trans_for_each_iter(trans, iter)
 992                 if (btree_iter_live(trans, iter) &&
 993                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT))
 994                         bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 995 out:
 996         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 997
 998         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 999                 percpu_ref_put(&trans->c->writes);
1000 out_reset:
1001         if (!ret)
1002                 reset_flags |= TRANS_RESET_NOTRAVERSE;
1003         if (!ret && (trans->flags & BTREE_INSERT_NOUNLOCK))
1004                 reset_flags |= TRANS_RESET_NOUNLOCK;
1005         bch2_trans_reset(trans, reset_flags);
1006
1007         return ret;
1008 err:
1009         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1010         if (ret)
1011                 goto out;
1012
1013         goto retry;
1014 }
1015
1016 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1017                       struct bkey_i *k, enum btree_update_flags flags)
1018 {
1019         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
1020                 .flags          = flags,
1021                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
1022                 .btree_id       = iter->btree_id,
1023                 .level          = iter->level,
1024                 .iter           = iter,
1025                 .k              = k
1026         };
1027         bool is_extent = (iter->flags & BTREE_ITER_IS_EXTENTS) != 0;
1028         int ret = 0;
1029
1030         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1031         BUG_ON(!iter->should_be_locked);
1032
1033 #ifdef CONFIG_BCACHEFS_DEBUG
1034         trans_for_each_update(trans, i)
1035                 BUG_ON(i != trans->updates &&
1036                        btree_insert_entry_cmp(i - 1, i) >= 0);
1037 #endif
1038
1039         if (is_extent) {
1040                 ret = extent_handle_overwrites(trans, &n);
1041                 if (ret)
1042                         return ret;
1043
1044                 iter->pos_after_commit = k->k.p;
1045                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
1046
1047                 if (bkey_deleted(&n.k->k))
1048                         return 0;
1049
1050                 n.iter = bch2_trans_get_iter(trans, n.btree_id, n.k->k.p,
1051                                              BTREE_ITER_INTENT|
1052                                              BTREE_ITER_NOT_EXTENTS);
1053                 bch2_trans_iter_put(trans, n.iter);
1054         }
1055
1056         BUG_ON(n.iter->flags & BTREE_ITER_IS_EXTENTS);
1057
1058         n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1059
1060         /*
1061          * Pending updates are kept sorted: first, find position of new update,
1062          * then delete/trim any updates the new update overwrites:
1063          */
1064         trans_for_each_update(trans, i)
1065                 if (btree_insert_entry_cmp(&n, i) <= 0)
1066                         break;
1067
1068         if (i < trans->updates + trans->nr_updates &&
1069             !btree_insert_entry_cmp(&n, i)) {
1070                 BUG_ON(i->trans_triggers_run);
1071                 *i = n;
1072         } else
1073                 array_insert_item(trans->updates, trans->nr_updates,
1074                                   i - trans->updates, n);
1075
1076         return 0;
1077 }
1078
1079 void bch2_trans_commit_hook(struct btree_trans *trans,
1080                             struct btree_trans_commit_hook *h)
1081 {
1082         h->next = trans->hooks;
1083         trans->hooks = h;
1084 }
1085
1086 int __bch2_btree_insert(struct btree_trans *trans,
1087                         enum btree_id id, struct bkey_i *k)
1088 {
1089         struct btree_iter *iter;
1090         int ret;
1091
1092         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1093                                    BTREE_ITER_INTENT);
1094
1095         ret   = bch2_btree_iter_traverse(iter) ?:
1096                 bch2_trans_update(trans, iter, k, 0);
1097         bch2_trans_iter_put(trans, iter);
1098         return ret;
1099 }
1100
1101 /**
1102  * bch2_btree_insert - insert keys into the extent btree
1103  * @c:                  pointer to struct bch_fs
1104  * @id:                 btree to insert into
1105  * @insert_keys:        list of keys to insert
1106  * @hook:               insert callback
1107  */
1108 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1109                       struct bkey_i *k,
1110                       struct disk_reservation *disk_res,
1111                       u64 *journal_seq, int flags)
1112 {
1113         return bch2_trans_do(c, disk_res, journal_seq, flags,
1114                              __bch2_btree_insert(&trans, id, k));
1115 }
1116
1117 int bch2_btree_delete_at(struct btree_trans *trans,
1118                          struct btree_iter *iter, unsigned update_flags)
1119 {
1120         struct bkey_i *k;
1121
1122         k = bch2_trans_kmalloc(trans, sizeof(*k));
1123         if (IS_ERR(k))
1124                 return PTR_ERR(k);
1125
1126         bkey_init(&k->k);
1127         k->k.p = iter->pos;
1128         return bch2_trans_update(trans, iter, k, update_flags);
1129 }
1130
1131 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1132                                   struct bpos start, struct bpos end,
1133                                   u64 *journal_seq)
1134 {
1135         struct btree_iter *iter;
1136         struct bkey_s_c k;
1137         int ret = 0;
1138
1139         iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
1140 retry:
1141         while ((bch2_trans_begin(trans),
1142                (k = bch2_btree_iter_peek(iter)).k) &&
1143                !(ret = bkey_err(k)) &&
1144                bkey_cmp(iter->pos, end) < 0) {
1145                 struct bkey_i delete;
1146
1147                 bkey_init(&delete.k);
1148
1149                 /*
1150                  * This could probably be more efficient for extents:
1151                  */
1152
1153                 /*
1154                  * For extents, iter.pos won't necessarily be the same as
1155                  * bkey_start_pos(k.k) (for non extents they always will be the
1156                  * same). It's important that we delete starting from iter.pos
1157                  * because the range we want to delete could start in the middle
1158                  * of k.
1159                  *
1160                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1161                  * bkey_start_pos(k.k)).
1162                  */
1163                 delete.k.p = iter->pos;
1164
1165                 if (btree_node_type_is_extents(iter->btree_id)) {
1166                         unsigned max_sectors =
1167                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1168
1169                         /* create the biggest key we can */
1170                         bch2_key_resize(&delete.k, max_sectors);
1171                         bch2_cut_back(end, &delete);
1172
1173                         ret = bch2_extent_trim_atomic(&delete, iter);
1174                         if (ret)
1175                                 break;
1176                 }
1177
1178                 ret   = bch2_trans_update(trans, iter, &delete, 0) ?:
1179                         bch2_trans_commit(trans, NULL, journal_seq,
1180                                         BTREE_INSERT_NOFAIL);
1181                 if (ret)
1182                         break;
1183
1184                 bch2_trans_cond_resched(trans);
1185         }
1186
1187         if (ret == -EINTR) {
1188                 ret = 0;
1189                 goto retry;
1190         }
1191
1192         bch2_trans_iter_free(trans, iter);
1193         return ret;
1194 }
1195
1196 /*
1197  * bch_btree_delete_range - delete everything within a given range
1198  *
1199  * Range is a half open interval - [start, end)
1200  */
1201 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1202                             struct bpos start, struct bpos end,
1203                             u64 *journal_seq)
1204 {
1205         return bch2_trans_do(c, NULL, journal_seq, 0,
1206                              bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
1207 }