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_trans_mark_gc(struct btree_trans *trans)
 352 {
 353         struct bch_fs *c = trans->c;
 354         struct btree_insert_entry *i;
 355
 356         trans_for_each_update(trans, i) {
 357                 /*
 358                  * XXX: synchronization of cached update triggers with gc
 359                  */
 360                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 361
 362                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 363                         bch2_mark_update(trans, i->iter, i->k,
 364                                          i->flags|BTREE_TRIGGER_GC);
 365         }
 366 }
 367
 368 static inline int
 369 bch2_trans_commit_write_locked(struct btree_trans *trans,
 370                                struct btree_insert_entry **stopped_at,
 371                                unsigned long trace_ip)
 372 {
 373         struct bch_fs *c = trans->c;
 374         struct btree_insert_entry *i;
 375         struct btree_trans_commit_hook *h;
 376         unsigned u64s = 0;
 377         bool marking = false;
 378         int ret;
 379
 380         if (race_fault()) {
 381                 trace_trans_restart_fault_inject(trans->ip, trace_ip);
 382                 trans->restarted = true;
 383                 return -EINTR;
 384         }
 385
 386         /*
 387          * Check if the insert will fit in the leaf node with the write lock
 388          * held, otherwise another thread could write the node changing the
 389          * amount of space available:
 390          */
 391
 392         prefetch(&trans->c->journal.flags);
 393
 394         h = trans->hooks;
 395         while (h) {
 396                 ret = h->fn(trans, h);
 397                 if (ret)
 398                         return ret;
 399                 h = h->next;
 400         }
 401
 402         trans_for_each_update(trans, i) {
 403                 /* Multiple inserts might go to same leaf: */
 404                 if (!same_leaf_as_prev(trans, i))
 405                         u64s = 0;
 406
 407                 u64s += i->k->k.u64s;
 408                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 409                         ? btree_key_can_insert(trans, i->iter, u64s)
 410                         : btree_key_can_insert_cached(trans, i->iter, u64s);
 411                 if (ret) {
 412                         *stopped_at = i;
 413                         return ret;
 414                 }
 415
 416                 if (btree_node_type_needs_gc(i->bkey_type))
 417                         marking = true;
 418         }
 419
 420         if (marking) {
 421                 percpu_down_read(&c->mark_lock);
 422         }
 423
 424         /* Must be called under mark_lock: */
 425         if (marking && trans->fs_usage_deltas &&
 426             !bch2_replicas_delta_list_marked(c, trans->fs_usage_deltas)) {
 427                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 428                 goto err;
 429         }
 430
 431         /*
 432          * Don't get journal reservation until after we know insert will
 433          * succeed:
 434          */
 435         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 436                 ret = bch2_trans_journal_res_get(trans,
 437                                 JOURNAL_RES_GET_NONBLOCK);
 438                 if (ret)
 439                         goto err;
 440         } else {
 441                 trans->journal_res.seq = c->journal.replay_journal_seq;
 442         }
 443
 444         if (unlikely(trans->extra_journal_entry_u64s)) {
 445                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 446                                   trans->extra_journal_entries,
 447                                   trans->extra_journal_entry_u64s);
 448
 449                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 450                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 451         }
 452
 453         /*
 454          * Not allowed to fail after we've gotten our journal reservation - we
 455          * have to use it:
 456          */
 457
 458         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 459                 if (bch2_journal_seq_verify)
 460                         trans_for_each_update(trans, i)
 461                                 i->k->k.version.lo = trans->journal_res.seq;
 462                 else if (bch2_inject_invalid_keys)
 463                         trans_for_each_update(trans, i)
 464                                 i->k->k.version = MAX_VERSION;
 465         }
 466
 467         trans_for_each_update(trans, i)
 468                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 469                         bch2_mark_update(trans, i->iter, i->k,
 470                                          i->flags);
 471
 472         if (marking && trans->fs_usage_deltas)
 473                 bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas);
 474
 475         if (unlikely(c->gc_pos.phase))
 476                 bch2_trans_mark_gc(trans);
 477
 478         trans_for_each_update(trans, i)
 479                 do_btree_insert_one(trans, i);
 480 err:
 481         if (marking) {
 482                 percpu_up_read(&c->mark_lock);
 483         }
 484
 485         return ret;
 486 }
 487
 488 static noinline int maybe_do_btree_merge(struct btree_trans *trans, struct btree_iter *iter)
 489 {
 490         struct btree_insert_entry *i;
 491         struct btree *b = iter_l(iter)->b;
 492         struct bkey_s_c old;
 493         int u64s_delta = 0;
 494         int ret;
 495
 496         /*
 497          * Inserting directly into interior nodes is an uncommon operation with
 498          * various weird edge cases: also, a lot of things about
 499          * BTREE_ITER_NODES iters need to be audited
 500          */
 501         if (unlikely(btree_iter_type(iter) != BTREE_ITER_KEYS))
 502                 return 0;
 503
 504         BUG_ON(iter->level);
 505
 506         trans_for_each_update(trans, i) {
 507                 if (iter_l(i->iter)->b != b)
 508                         continue;
 509
 510                 old = bch2_btree_iter_peek_slot(i->iter);
 511                 ret = bkey_err(old);
 512                 if (ret)
 513                         return ret;
 514
 515                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 516                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 517         }
 518
 519         if (u64s_delta > 0)
 520                 return 0;
 521
 522         return bch2_foreground_maybe_merge(trans, iter,
 523                                 iter->level, trans->flags);
 524 }
 525
 526 /*
 527  * Get journal reservation, take write locks, and attempt to do btree update(s):
 528  */
 529 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 530                                        struct btree_insert_entry **stopped_at,
 531                                        unsigned long trace_ip)
 532 {
 533         struct bch_fs *c = trans->c;
 534         struct btree_insert_entry *i;
 535         struct btree_iter *iter;
 536         int ret;
 537
 538         trans_for_each_update(trans, i) {
 539                 struct btree *b;
 540
 541                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 542
 543                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED)
 544                         continue;
 545
 546                 b = iter_l(i->iter)->b;
 547                 if (b->sib_u64s[0] < c->btree_foreground_merge_threshold ||
 548                     b->sib_u64s[1] < c->btree_foreground_merge_threshold) {
 549                         ret = maybe_do_btree_merge(trans, i->iter);
 550                         if (unlikely(ret))
 551                                 return ret;
 552                 }
 553         }
 554
 555         trans_for_each_update(trans, i)
 556                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 557
 558         ret = bch2_journal_preres_get(&c->journal,
 559                         &trans->journal_preres, trans->journal_preres_u64s,
 560                         JOURNAL_RES_GET_NONBLOCK|
 561                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 562                          ? JOURNAL_RES_GET_RESERVED : 0));
 563         if (unlikely(ret == -EAGAIN))
 564                 ret = bch2_trans_journal_preres_get_cold(trans,
 565                                                 trans->journal_preres_u64s, trace_ip);
 566         if (unlikely(ret))
 567                 return ret;
 568
 569         /*
 570          * Can't be holding any read locks when we go to take write locks:
 571          * another thread could be holding an intent lock on the same node we
 572          * have a read lock on, and it'll block trying to take a write lock
 573          * (because we hold a read lock) and it could be blocking us by holding
 574          * its own read lock (while we're trying to to take write locks).
 575          *
 576          * note - this must be done after bch2_trans_journal_preres_get_cold()
 577          * or anything else that might call bch2_trans_relock(), since that
 578          * would just retake the read locks:
 579          */
 580         trans_for_each_iter(trans, iter)
 581                 if (iter->nodes_locked != iter->nodes_intent_locked &&
 582                     !bch2_btree_iter_upgrade(iter, 1)) {
 583                         trace_trans_restart_upgrade(trans->ip, trace_ip,
 584                                                     iter->btree_id,
 585                                                     &iter->real_pos);
 586                         trans->restarted = true;
 587                         return -EINTR;
 588                 }
 589
 590         trans_for_each_update(trans, i) {
 591                 const char *invalid = bch2_bkey_invalid(c,
 592                                 bkey_i_to_s_c(i->k), i->bkey_type);
 593                 if (invalid) {
 594                         char buf[200];
 595
 596                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 597                         bch_err(c, "invalid bkey %s on insert: %s\n", buf, invalid);
 598                         bch2_fatal_error(c);
 599                 }
 600                 btree_insert_entry_checks(trans, i);
 601         }
 602         bch2_btree_trans_verify_locks(trans);
 603
 604         trans_for_each_update(trans, i)
 605                 if (!same_leaf_as_prev(trans, i))
 606                         bch2_btree_node_lock_for_insert(trans, i->iter,
 607                                         iter_l(i->iter)->b);
 608
 609         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 610
 611         trans_for_each_update(trans, i)
 612                 if (!same_leaf_as_prev(trans, i))
 613                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 614                                                              i->iter);
 615
 616         if (!ret && trans->journal_pin)
 617                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 618                                      trans->journal_pin, NULL);
 619
 620         /*
 621          * Drop journal reservation after dropping write locks, since dropping
 622          * the journal reservation may kick off a journal write:
 623          */
 624         bch2_journal_res_put(&c->journal, &trans->journal_res);
 625
 626         if (unlikely(ret))
 627                 return ret;
 628
 629         bch2_trans_downgrade(trans);
 630
 631         return 0;
 632 }
 633
 634 static int journal_reclaim_wait_done(struct bch_fs *c)
 635 {
 636         int ret = bch2_journal_error(&c->journal) ?:
 637                 !bch2_btree_key_cache_must_wait(c);
 638
 639         if (!ret)
 640                 journal_reclaim_kick(&c->journal);
 641         return ret;
 642 }
 643
 644 static noinline
 645 int bch2_trans_commit_error(struct btree_trans *trans,
 646                             struct btree_insert_entry *i,
 647                             int ret, unsigned long trace_ip)
 648 {
 649         struct bch_fs *c = trans->c;
 650
 651         switch (ret) {
 652         case BTREE_INSERT_BTREE_NODE_FULL:
 653                 ret = bch2_btree_split_leaf(trans, i->iter, trans->flags);
 654                 if (!ret)
 655                         return 0;
 656
 657                 if (ret == -EINTR)
 658                         trace_trans_restart_btree_node_split(trans->ip, trace_ip,
 659                                                              i->iter->btree_id,
 660                                                              &i->iter->real_pos);
 661                 break;
 662         case BTREE_INSERT_NEED_MARK_REPLICAS:
 663                 bch2_trans_unlock(trans);
 664
 665                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 666                 if (ret)
 667                         break;
 668
 669                 if (bch2_trans_relock(trans))
 670                         return 0;
 671
 672                 trace_trans_restart_mark_replicas(trans->ip, trace_ip);
 673                 ret = -EINTR;
 674                 break;
 675         case BTREE_INSERT_NEED_JOURNAL_RES:
 676                 bch2_trans_unlock(trans);
 677
 678                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 679                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED)) {
 680                         trans->restarted = true;
 681                         ret = -EAGAIN;
 682                         break;
 683                 }
 684
 685                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 686                 if (ret)
 687                         break;
 688
 689                 if (bch2_trans_relock(trans))
 690                         return 0;
 691
 692                 trace_trans_restart_journal_res_get(trans->ip, trace_ip);
 693                 ret = -EINTR;
 694                 break;
 695         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 696                 bch2_trans_unlock(trans);
 697
 698                 trace_trans_blocked_journal_reclaim(trans->ip, trace_ip);
 699
 700                 wait_event_freezable(c->journal.reclaim_wait,
 701                                      (ret = journal_reclaim_wait_done(c)));
 702                 if (ret < 0)
 703                         break;
 704
 705                 if (bch2_trans_relock(trans))
 706                         return 0;
 707
 708                 trace_trans_restart_journal_reclaim(trans->ip, trace_ip);
 709                 ret = -EINTR;
 710                 break;
 711         default:
 712                 BUG_ON(ret >= 0);
 713                 break;
 714         }
 715
 716         BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
 717         BUG_ON(ret == -ENOSPC && (trans->flags & BTREE_INSERT_NOFAIL));
 718
 719         return ret;
 720 }
 721
 722 static noinline int
 723 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 724 {
 725         struct bch_fs *c = trans->c;
 726         int ret;
 727
 728         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 729                 return -EROFS;
 730
 731         bch2_trans_unlock(trans);
 732
 733         ret = bch2_fs_read_write_early(c);
 734         if (ret)
 735                 return ret;
 736
 737         percpu_ref_get(&c->writes);
 738         return 0;
 739 }
 740
 741 static int extent_handle_overwrites(struct btree_trans *trans,
 742                                     struct btree_insert_entry *i)
 743 {
 744         struct bch_fs *c = trans->c;
 745         struct btree_iter *iter, *update_iter;
 746         struct bpos start = bkey_start_pos(&i->k->k);
 747         struct bkey_i *update;
 748         struct bkey_s_c k;
 749         int ret = 0, compressed_sectors;
 750
 751         iter = bch2_trans_get_iter(trans, i->btree_id, start,
 752                                    BTREE_ITER_INTENT|
 753                                    BTREE_ITER_WITH_UPDATES|
 754                                    BTREE_ITER_NOT_EXTENTS);
 755         k = bch2_btree_iter_peek(iter);
 756         if (!k.k || (ret = bkey_err(k)))
 757                 goto out;
 758
 759         if (bch2_bkey_maybe_mergable(k.k, &i->k->k)) {
 760                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 761                 if ((ret = PTR_ERR_OR_ZERO(update)))
 762                         goto out;
 763
 764                 bkey_reassemble(update, k);
 765
 766                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(i->k))) {
 767                         update_iter = bch2_trans_copy_iter(trans, iter);
 768                         ret = bch2_btree_delete_at(trans, update_iter, i->flags);
 769                         bch2_trans_iter_put(trans, update_iter);
 770
 771                         if (ret)
 772                                 goto out;
 773
 774                         i->k = update;
 775                         goto next;
 776                 }
 777         }
 778
 779         if (!bkey_cmp(k.k->p, bkey_start_pos(&i->k->k)))
 780                 goto next;
 781
 782         while (bkey_cmp(i->k->k.p, bkey_start_pos(k.k)) > 0) {
 783                 /*
 784                  * If we're going to be splitting a compressed extent, note it
 785                  * so that __bch2_trans_commit() can increase our disk
 786                  * reservation:
 787                  */
 788                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0 &&
 789                     bkey_cmp(k.k->p, i->k->k.p) > 0 &&
 790                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
 791                         trans->extra_journal_res += compressed_sectors;
 792
 793                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 794                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 795                         if ((ret = PTR_ERR_OR_ZERO(update)))
 796                                 goto out;
 797
 798                         bkey_reassemble(update, k);
 799
 800                         bch2_cut_back(start, update);
 801
 802                         update_iter = bch2_trans_get_iter(trans, i->btree_id, update->k.p,
 803                                                           BTREE_ITER_NOT_EXTENTS|
 804                                                           BTREE_ITER_INTENT);
 805                         ret = bch2_btree_iter_traverse(update_iter);
 806                         if (ret) {
 807                                 bch2_trans_iter_put(trans, update_iter);
 808                                 goto out;
 809                         }
 810
 811                         bch2_trans_update(trans, update_iter, update,
 812                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 813                                           i->flags);
 814                         bch2_trans_iter_put(trans, update_iter);
 815                 }
 816
 817                 if (bkey_cmp(k.k->p, i->k->k.p) <= 0) {
 818                         update_iter = bch2_trans_copy_iter(trans, iter);
 819                         ret = bch2_btree_delete_at(trans, update_iter,
 820                                                    i->flags);
 821                         bch2_trans_iter_put(trans, update_iter);
 822
 823                         if (ret)
 824                                 goto out;
 825                 }
 826
 827                 if (bkey_cmp(k.k->p, i->k->k.p) > 0) {
 828                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 829                         if ((ret = PTR_ERR_OR_ZERO(update)))
 830                                 goto out;
 831
 832                         bkey_reassemble(update, k);
 833                         bch2_cut_front(i->k->k.p, update);
 834
 835                         bch2_trans_update(trans, iter, update, i->flags);
 836                         goto out;
 837                 }
 838 next:
 839                 k = bch2_btree_iter_next(iter);
 840                 if (!k.k || (ret = bkey_err(k)))
 841                         goto out;
 842         }
 843
 844         bch2_bkey_merge(c, bkey_i_to_s(i->k), k);
 845 out:
 846         bch2_trans_iter_put(trans, iter);
 847
 848         return ret;
 849 }
 850
 851 int __bch2_trans_commit(struct btree_trans *trans)
 852 {
 853         struct btree_insert_entry *i = NULL;
 854         struct btree_iter *iter;
 855         bool trans_trigger_run;
 856         unsigned u64s;
 857         int ret = 0;
 858
 859         if (!trans->nr_updates &&
 860             !trans->extra_journal_entry_u64s)
 861                 goto out_reset;
 862
 863         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 864                 lockdep_assert_held(&trans->c->gc_lock);
 865
 866         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 867
 868         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 869         trans->journal_preres_u64s      = 0;
 870
 871         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 872             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 873                 ret = bch2_trans_commit_get_rw_cold(trans);
 874                 if (ret)
 875                         goto out_reset;
 876         }
 877
 878 #ifdef CONFIG_BCACHEFS_DEBUG
 879         trans_for_each_update(trans, i)
 880                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 881                     !(i->flags & BTREE_TRIGGER_NORUN))
 882                         bch2_btree_key_cache_verify_clean(trans,
 883                                         i->btree_id, i->k->k.p);
 884 #endif
 885
 886         /*
 887          * Running triggers will append more updates to the list of updates as
 888          * we're walking it:
 889          */
 890         do {
 891                 trans_trigger_run = false;
 892
 893                 trans_for_each_update(trans, i) {
 894                         if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 895                             !i->trans_triggers_run) {
 896                                 i->trans_triggers_run = true;
 897                                 trans_trigger_run = true;
 898
 899                                 ret = bch2_trans_mark_update(trans, i->iter,
 900                                                              i->k, i->flags);
 901                                 if (unlikely(ret)) {
 902                                         if (ret == -EINTR)
 903                                                 trace_trans_restart_mark(trans->ip, _RET_IP_,
 904                                                                          i->iter->btree_id,
 905                                                                          &i->iter->pos);
 906                                         goto out;
 907                                 }
 908                         }
 909                 }
 910         } while (trans_trigger_run);
 911
 912         trans_for_each_update(trans, i) {
 913                 BUG_ON(!i->iter->should_be_locked);
 914
 915                 if (unlikely(!bch2_btree_iter_upgrade(i->iter, i->level + 1))) {
 916                         trace_trans_restart_upgrade(trans->ip, _RET_IP_,
 917                                                     i->iter->btree_id,
 918                                                     &i->iter->pos);
 919                         trans->restarted = true;
 920                         ret = -EINTR;
 921                         goto out;
 922                 }
 923
 924                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 925
 926                 u64s = jset_u64s(i->k->k.u64s);
 927                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 928                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 929                         trans->journal_preres_u64s += u64s;
 930                 trans->journal_u64s += u64s;
 931         }
 932
 933         if (trans->extra_journal_res) {
 934                 ret = bch2_disk_reservation_add(trans->c, trans->disk_res,
 935                                 trans->extra_journal_res,
 936                                 (trans->flags & BTREE_INSERT_NOFAIL)
 937                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 938                 if (ret)
 939                         goto err;
 940         }
 941 retry:
 942         BUG_ON(trans->restarted);
 943         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 944
 945         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
 946
 947         /* make sure we didn't drop or screw up locks: */
 948         bch2_btree_trans_verify_locks(trans);
 949
 950         if (ret)
 951                 goto err;
 952
 953         trans_for_each_iter(trans, iter)
 954                 if (btree_iter_live(trans, iter) &&
 955                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT))
 956                         bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 957 out:
 958         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 959
 960         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 961                 percpu_ref_put(&trans->c->writes);
 962 out_reset:
 963         trans->extra_journal_res        = 0;
 964         trans->nr_updates               = 0;
 965         trans->hooks                    = NULL;
 966         trans->extra_journal_entries    = NULL;
 967         trans->extra_journal_entry_u64s = 0;
 968
 969         if (trans->fs_usage_deltas) {
 970                 trans->fs_usage_deltas->used = 0;
 971                 memset(&trans->fs_usage_deltas->memset_start, 0,
 972                        (void *) &trans->fs_usage_deltas->memset_end -
 973                        (void *) &trans->fs_usage_deltas->memset_start);
 974         }
 975
 976         return ret;
 977 err:
 978         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
 979         if (ret)
 980                 goto out;
 981
 982         goto retry;
 983 }
 984
 985 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 986                       struct bkey_i *k, enum btree_update_flags flags)
 987 {
 988         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 989                 .flags          = flags,
 990                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
 991                 .btree_id       = iter->btree_id,
 992                 .level          = iter->level,
 993                 .iter           = iter,
 994                 .k              = k
 995         };
 996         bool is_extent = (iter->flags & BTREE_ITER_IS_EXTENTS) != 0;
 997         int ret = 0;
 998
 999         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1000         BUG_ON(!iter->should_be_locked);
1001
1002 #ifdef CONFIG_BCACHEFS_DEBUG
1003         trans_for_each_update(trans, i)
1004                 BUG_ON(i != trans->updates &&
1005                        btree_insert_entry_cmp(i - 1, i) >= 0);
1006 #endif
1007
1008         if (is_extent) {
1009                 ret = extent_handle_overwrites(trans, &n);
1010                 if (ret)
1011                         return ret;
1012
1013                 iter->pos_after_commit = k->k.p;
1014                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
1015
1016                 if (bkey_deleted(&n.k->k))
1017                         return 0;
1018
1019                 n.iter = bch2_trans_get_iter(trans, n.btree_id, n.k->k.p,
1020                                              BTREE_ITER_INTENT|
1021                                              BTREE_ITER_NOT_EXTENTS);
1022                 ret = bch2_btree_iter_traverse(n.iter);
1023                 bch2_trans_iter_put(trans, n.iter);
1024
1025                 if (ret)
1026                         return ret;
1027         }
1028
1029         BUG_ON(n.iter->flags & BTREE_ITER_IS_EXTENTS);
1030
1031         n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1032
1033         /*
1034          * Pending updates are kept sorted: first, find position of new update,
1035          * then delete/trim any updates the new update overwrites:
1036          */
1037         trans_for_each_update(trans, i)
1038                 if (btree_insert_entry_cmp(&n, i) <= 0)
1039                         break;
1040
1041         if (i < trans->updates + trans->nr_updates &&
1042             !btree_insert_entry_cmp(&n, i)) {
1043                 BUG_ON(i->trans_triggers_run);
1044                 *i = n;
1045         } else
1046                 array_insert_item(trans->updates, trans->nr_updates,
1047                                   i - trans->updates, n);
1048
1049         return 0;
1050 }
1051
1052 void bch2_trans_commit_hook(struct btree_trans *trans,
1053                             struct btree_trans_commit_hook *h)
1054 {
1055         h->next = trans->hooks;
1056         trans->hooks = h;
1057 }
1058
1059 int __bch2_btree_insert(struct btree_trans *trans,
1060                         enum btree_id id, struct bkey_i *k)
1061 {
1062         struct btree_iter *iter;
1063         int ret;
1064
1065         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1066                                    BTREE_ITER_INTENT);
1067
1068         ret   = bch2_btree_iter_traverse(iter) ?:
1069                 bch2_trans_update(trans, iter, k, 0);
1070         bch2_trans_iter_put(trans, iter);
1071         return ret;
1072 }
1073
1074 /**
1075  * bch2_btree_insert - insert keys into the extent btree
1076  * @c:                  pointer to struct bch_fs
1077  * @id:                 btree to insert into
1078  * @insert_keys:        list of keys to insert
1079  * @hook:               insert callback
1080  */
1081 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1082                       struct bkey_i *k,
1083                       struct disk_reservation *disk_res,
1084                       u64 *journal_seq, int flags)
1085 {
1086         return bch2_trans_do(c, disk_res, journal_seq, flags,
1087                              __bch2_btree_insert(&trans, id, k));
1088 }
1089
1090 int bch2_btree_delete_at(struct btree_trans *trans,
1091                          struct btree_iter *iter, unsigned update_flags)
1092 {
1093         struct bkey_i *k;
1094
1095         k = bch2_trans_kmalloc(trans, sizeof(*k));
1096         if (IS_ERR(k))
1097                 return PTR_ERR(k);
1098
1099         bkey_init(&k->k);
1100         k->k.p = iter->pos;
1101         return bch2_trans_update(trans, iter, k, update_flags);
1102 }
1103
1104 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1105                                   struct bpos start, struct bpos end,
1106                                   u64 *journal_seq)
1107 {
1108         struct btree_iter *iter;
1109         struct bkey_s_c k;
1110         int ret = 0;
1111
1112         iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
1113 retry:
1114         while ((bch2_trans_begin(trans),
1115                (k = bch2_btree_iter_peek(iter)).k) &&
1116                !(ret = bkey_err(k)) &&
1117                bkey_cmp(iter->pos, end) < 0) {
1118                 struct bkey_i delete;
1119
1120                 bkey_init(&delete.k);
1121
1122                 /*
1123                  * This could probably be more efficient for extents:
1124                  */
1125
1126                 /*
1127                  * For extents, iter.pos won't necessarily be the same as
1128                  * bkey_start_pos(k.k) (for non extents they always will be the
1129                  * same). It's important that we delete starting from iter.pos
1130                  * because the range we want to delete could start in the middle
1131                  * of k.
1132                  *
1133                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1134                  * bkey_start_pos(k.k)).
1135                  */
1136                 delete.k.p = iter->pos;
1137
1138                 if (btree_node_type_is_extents(iter->btree_id)) {
1139                         unsigned max_sectors =
1140                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1141
1142                         /* create the biggest key we can */
1143                         bch2_key_resize(&delete.k, max_sectors);
1144                         bch2_cut_back(end, &delete);
1145
1146                         ret = bch2_extent_trim_atomic(&delete, iter);
1147                         if (ret)
1148                                 break;
1149                 }
1150
1151                 ret   = bch2_trans_update(trans, iter, &delete, 0) ?:
1152                         bch2_trans_commit(trans, NULL, journal_seq,
1153                                         BTREE_INSERT_NOFAIL);
1154                 if (ret)
1155                         break;
1156
1157                 bch2_trans_cond_resched(trans);
1158         }
1159
1160         if (ret == -EINTR) {
1161                 ret = 0;
1162                 goto retry;
1163         }
1164
1165         bch2_trans_iter_free(trans, iter);
1166         return ret;
1167 }
1168
1169 /*
1170  * bch_btree_delete_range - delete everything within a given range
1171  *
1172  * Range is a half open interval - [start, end)
1173  */
1174 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1175                             struct bpos start, struct bpos end,
1176                             u64 *journal_seq)
1177 {
1178         return bch2_trans_do(c, NULL, journal_seq, 0,
1179                              bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
1180 }