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 "errcode.h"
  14 #include "error.h"
  15 #include "extent_update.h"
  16 #include "journal.h"
  17 #include "journal_reclaim.h"
  18 #include "keylist.h"
  19 #include "recovery.h"
  20 #include "subvolume.h"
  21 #include "replicas.h"
  22
  23 #include <linux/prefetch.h>
  24 #include <linux/sort.h>
  25 #include <trace/events/bcachefs.h>
  26
  27 static int __must_check
  28 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  29                           struct bkey_i *, enum btree_update_flags);
  30
  31 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  32                                          const struct btree_insert_entry *r)
  33 {
  34         return   cmp_int(l->btree_id,   r->btree_id) ?:
  35                  cmp_int(l->cached,     r->cached) ?:
  36                  -cmp_int(l->level,     r->level) ?:
  37                  bpos_cmp(l->k->k.p,    r->k->k.p);
  38 }
  39
  40 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  41 {
  42         return i->path->l + i->level;
  43 }
  44
  45 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  46                                      struct btree_insert_entry *i)
  47 {
  48         return i != trans->updates &&
  49                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  50 }
  51
  52 static inline bool same_leaf_as_next(struct btree_trans *trans,
  53                                      struct btree_insert_entry *i)
  54 {
  55         return i + 1 < trans->updates + trans->nr_updates &&
  56                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  57 }
  58
  59 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  60                                            struct btree_path *path,
  61                                            struct btree *b)
  62 {
  63         struct bch_fs *c = trans->c;
  64
  65         if (path->cached)
  66                 return;
  67
  68         if (unlikely(btree_node_just_written(b)) &&
  69             bch2_btree_post_write_cleanup(c, b))
  70                 bch2_trans_node_reinit_iter(trans, b);
  71
  72         /*
  73          * If the last bset has been written, or if it's gotten too big - start
  74          * a new bset to insert into:
  75          */
  76         if (want_new_bset(c, b))
  77                 bch2_btree_init_next(trans, b);
  78 }
  79
  80 /* Inserting into a given leaf node (last stage of insert): */
  81
  82 /* Handle overwrites and do insert, for non extents: */
  83 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  84                                 struct btree_path *path,
  85                                 struct btree *b,
  86                                 struct btree_node_iter *node_iter,
  87                                 struct bkey_i *insert)
  88 {
  89         struct bkey_packed *k;
  90         unsigned clobber_u64s = 0, new_u64s = 0;
  91
  92         EBUG_ON(btree_node_just_written(b));
  93         EBUG_ON(bset_written(b, btree_bset_last(b)));
  94         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  95         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
  96         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
  97         EBUG_ON(insert->k.u64s >
  98                 bch_btree_keys_u64s_remaining(trans->c, b));
  99
 100         k = bch2_btree_node_iter_peek_all(node_iter, b);
 101         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 102                 k = NULL;
 103
 104         /* @k is the key being overwritten/deleted, if any: */
 105         EBUG_ON(k && bkey_deleted(k));
 106
 107         /* Deleting, but not found? nothing to do: */
 108         if (bkey_deleted(&insert->k) && !k)
 109                 return false;
 110
 111         if (bkey_deleted(&insert->k)) {
 112                 /* Deleting: */
 113                 btree_account_key_drop(b, k);
 114                 k->type = KEY_TYPE_deleted;
 115
 116                 if (k->needs_whiteout)
 117                         push_whiteout(trans->c, b, insert->k.p);
 118                 k->needs_whiteout = false;
 119
 120                 if (k >= btree_bset_last(b)->start) {
 121                         clobber_u64s = k->u64s;
 122                         bch2_bset_delete(b, k, clobber_u64s);
 123                         goto fix_iter;
 124                 } else {
 125                         bch2_btree_path_fix_key_modified(trans, b, k);
 126                 }
 127
 128                 return true;
 129         }
 130
 131         if (k) {
 132                 /* Overwriting: */
 133                 btree_account_key_drop(b, k);
 134                 k->type = KEY_TYPE_deleted;
 135
 136                 insert->k.needs_whiteout = k->needs_whiteout;
 137                 k->needs_whiteout = false;
 138
 139                 if (k >= btree_bset_last(b)->start) {
 140                         clobber_u64s = k->u64s;
 141                         goto overwrite;
 142                 } else {
 143                         bch2_btree_path_fix_key_modified(trans, b, k);
 144                 }
 145         }
 146
 147         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 148 overwrite:
 149         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 150         new_u64s = k->u64s;
 151 fix_iter:
 152         if (clobber_u64s != new_u64s)
 153                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 154                                          clobber_u64s, new_u64s);
 155         return true;
 156 }
 157
 158 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 159                                unsigned i, u64 seq)
 160 {
 161         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 162         struct btree_write *w = container_of(pin, struct btree_write, journal);
 163         struct btree *b = container_of(w, struct btree, writes[i]);
 164         struct btree_trans trans;
 165         unsigned long old, new, v;
 166         unsigned idx = w - b->writes;
 167
 168         bch2_trans_init(&trans, c, 0, 0);
 169
 170         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 171         v = READ_ONCE(b->flags);
 172
 173         do {
 174                 old = new = v;
 175
 176                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 177                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 178                     w->journal.seq != seq)
 179                         break;
 180
 181                 new |= 1 << BTREE_NODE_need_write;
 182         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 183
 184         btree_node_write_if_need(c, b, SIX_LOCK_read);
 185         six_unlock_read(&b->c.lock);
 186
 187         bch2_trans_exit(&trans);
 188         return 0;
 189 }
 190
 191 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 192 {
 193         return __btree_node_flush(j, pin, 0, seq);
 194 }
 195
 196 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 197 {
 198         return __btree_node_flush(j, pin, 1, seq);
 199 }
 200
 201 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 202                                        struct btree *b, u64 seq)
 203 {
 204         struct btree_write *w = btree_current_write(b);
 205
 206         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 207                              btree_node_write_idx(b) == 0
 208                              ? btree_node_flush0
 209                              : btree_node_flush1);
 210 }
 211
 212 /**
 213  * btree_insert_key - insert a key one key into a leaf node
 214  */
 215 static void btree_insert_key_leaf(struct btree_trans *trans,
 216                                   struct btree_insert_entry *insert)
 217 {
 218         struct bch_fs *c = trans->c;
 219         struct btree *b = insert_l(insert)->b;
 220         struct bset_tree *t = bset_tree_last(b);
 221         struct bset *i = bset(b, t);
 222         int old_u64s = bset_u64s(t);
 223         int old_live_u64s = b->nr.live_u64s;
 224         int live_u64s_added, u64s_added;
 225
 226         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 227                                         &insert_l(insert)->iter, insert->k)))
 228                 return;
 229
 230         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 231                                          le64_to_cpu(i->journal_seq)));
 232
 233         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 234
 235         if (unlikely(!btree_node_dirty(b)))
 236                 set_btree_node_dirty_acct(c, b);
 237
 238         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 239         u64s_added = (int) bset_u64s(t) - old_u64s;
 240
 241         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 242                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 243         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 244                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 245
 246         if (u64s_added > live_u64s_added &&
 247             bch2_maybe_compact_whiteouts(c, b))
 248                 bch2_trans_node_reinit_iter(trans, b);
 249 }
 250
 251 /* Cached btree updates: */
 252
 253 /* Normal update interface: */
 254
 255 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 256                                              struct btree_insert_entry *i)
 257 {
 258         BUG_ON(bpos_cmp(i->k->k.p, i->path->pos));
 259         BUG_ON(i->cached        != i->path->cached);
 260         BUG_ON(i->level         != i->path->level);
 261         BUG_ON(i->btree_id      != i->path->btree_id);
 262         EBUG_ON(!i->level &&
 263                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 264                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 265                 i->k->k.p.snapshot &&
 266                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 267 }
 268
 269 static noinline int
 270 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 271                                    unsigned long trace_ip)
 272 {
 273         struct bch_fs *c = trans->c;
 274         int ret;
 275
 276         bch2_trans_unlock(trans);
 277
 278         ret = bch2_journal_preres_get(&c->journal,
 279                         &trans->journal_preres, u64s, 0);
 280         if (ret)
 281                 return ret;
 282
 283         ret = bch2_trans_relock(trans);
 284         if (ret) {
 285                 trace_and_count(c, trans_restart_journal_preres_get, trans, trace_ip, 0);
 286                 return ret;
 287         }
 288
 289         return 0;
 290 }
 291
 292 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 293                                              unsigned flags)
 294 {
 295         struct bch_fs *c = trans->c;
 296         int ret;
 297
 298         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 299                                    trans->journal_u64s,
 300                                    flags|
 301                                    (trans->flags & JOURNAL_WATERMARK_MASK));
 302
 303         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 304 }
 305
 306 #define JSET_ENTRY_LOG_U64s             4
 307
 308 static void journal_transaction_name(struct btree_trans *trans)
 309 {
 310         struct bch_fs *c = trans->c;
 311         struct journal *j = &c->journal;
 312         struct jset_entry *entry =
 313                 bch2_journal_add_entry(j, &trans->journal_res,
 314                                        BCH_JSET_ENTRY_log, 0, 0,
 315                                        JSET_ENTRY_LOG_U64s);
 316         struct jset_entry_log *l =
 317                 container_of(entry, struct jset_entry_log, entry);
 318
 319         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 320 }
 321
 322 static inline enum btree_insert_ret
 323 btree_key_can_insert(struct btree_trans *trans,
 324                      struct btree *b,
 325                      unsigned u64s)
 326 {
 327         struct bch_fs *c = trans->c;
 328
 329         if (!bch2_btree_node_insert_fits(c, b, u64s))
 330                 return BTREE_INSERT_BTREE_NODE_FULL;
 331
 332         return BTREE_INSERT_OK;
 333 }
 334
 335 static enum btree_insert_ret
 336 btree_key_can_insert_cached(struct btree_trans *trans,
 337                             struct btree_path *path,
 338                             unsigned u64s)
 339 {
 340         struct bch_fs *c = trans->c;
 341         struct bkey_cached *ck = (void *) path->l[0].b;
 342         unsigned new_u64s;
 343         struct bkey_i *new_k;
 344
 345         EBUG_ON(path->level);
 346
 347         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 348             bch2_btree_key_cache_must_wait(c) &&
 349             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 350                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 351
 352         /*
 353          * bch2_varint_decode can read past the end of the buffer by at most 7
 354          * bytes (it won't be used):
 355          */
 356         u64s += 1;
 357
 358         if (u64s <= ck->u64s)
 359                 return BTREE_INSERT_OK;
 360
 361         new_u64s        = roundup_pow_of_two(u64s);
 362         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 363         if (!new_k) {
 364                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 365                         bch2_btree_ids[path->btree_id], new_u64s);
 366                 return -ENOMEM;
 367         }
 368
 369         ck->u64s        = new_u64s;
 370         ck->k           = new_k;
 371         return 0;
 372 }
 373
 374 /* Triggers: */
 375
 376 static int run_one_mem_trigger(struct btree_trans *trans,
 377                                struct btree_insert_entry *i,
 378                                unsigned flags)
 379 {
 380         struct bkey_s_c old = { &i->old_k, i->old_v };
 381         struct bkey_i *new = i->k;
 382         int ret;
 383
 384         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 385                 return 0;
 386
 387         if (!btree_node_type_needs_gc(i->btree_id))
 388                 return 0;
 389
 390         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 391             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 392             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 393                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 394                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 395         } else {
 396                 struct bkey             _deleted = KEY(0, 0, 0);
 397                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 398
 399                 _deleted.p = i->path->pos;
 400
 401                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 402                                 BTREE_TRIGGER_INSERT|flags) ?:
 403                         bch2_mark_key(trans, old, deleted,
 404                                 BTREE_TRIGGER_OVERWRITE|flags);
 405         }
 406
 407         return ret;
 408 }
 409
 410 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 411                                  bool overwrite)
 412 {
 413         /*
 414          * Transactional triggers create new btree_insert_entries, so we can't
 415          * pass them a pointer to a btree_insert_entry, that memory is going to
 416          * move:
 417          */
 418         struct bkey old_k = i->old_k;
 419         struct bkey_s_c old = { &old_k, i->old_v };
 420
 421         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 422             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 423                 return 0;
 424
 425         if (!i->insert_trigger_run &&
 426             !i->overwrite_trigger_run &&
 427             bch2_bkey_ops[old.k->type].trans_trigger ==
 428             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 429             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 430                 i->overwrite_trigger_run = true;
 431                 i->insert_trigger_run = true;
 432                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 433                                            BTREE_TRIGGER_INSERT|
 434                                            BTREE_TRIGGER_OVERWRITE|
 435                                            i->flags) ?: 1;
 436         } else if (overwrite && !i->overwrite_trigger_run) {
 437                 i->overwrite_trigger_run = true;
 438                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 439         } else if (!overwrite && !i->insert_trigger_run) {
 440                 i->insert_trigger_run = true;
 441                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 442         } else {
 443                 return 0;
 444         }
 445 }
 446
 447 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 448                               struct btree_insert_entry *btree_id_start)
 449 {
 450         struct btree_insert_entry *i;
 451         bool trans_trigger_run;
 452         int ret, overwrite;
 453
 454         for (overwrite = 1; overwrite >= 0; --overwrite) {
 455
 456                 /*
 457                  * Running triggers will append more updates to the list of updates as
 458                  * we're walking it:
 459                  */
 460                 do {
 461                         trans_trigger_run = false;
 462
 463                         for (i = btree_id_start;
 464                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 465                              i++) {
 466                                 if (i->btree_id != btree_id)
 467                                         continue;
 468
 469                                 ret = run_one_trans_trigger(trans, i, overwrite);
 470                                 if (ret < 0)
 471                                         return ret;
 472                                 if (ret)
 473                                         trans_trigger_run = true;
 474                         }
 475                 } while (trans_trigger_run);
 476         }
 477
 478         return 0;
 479 }
 480
 481 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 482 {
 483         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 484         unsigned btree_id = 0;
 485         int ret = 0;
 486
 487         /*
 488          *
 489          * For a given btree, this algorithm runs insert triggers before
 490          * overwrite triggers: this is so that when extents are being moved
 491          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 492          * they are re-added.
 493          */
 494         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 495                 if (btree_id == BTREE_ID_alloc)
 496                         continue;
 497
 498                 while (btree_id_start < trans->updates + trans->nr_updates &&
 499                        btree_id_start->btree_id < btree_id)
 500                         btree_id_start++;
 501
 502                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 503                 if (ret)
 504                         return ret;
 505         }
 506
 507         trans_for_each_update(trans, i) {
 508                 if (i->btree_id > BTREE_ID_alloc)
 509                         break;
 510                 if (i->btree_id == BTREE_ID_alloc) {
 511                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 512                         if (ret)
 513                                 return ret;
 514                         break;
 515                 }
 516         }
 517
 518         trans_for_each_update(trans, i)
 519                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 520                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 521                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 522
 523         return 0;
 524 }
 525
 526 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 527 {
 528         struct bch_fs *c = trans->c;
 529         struct btree_insert_entry *i;
 530         int ret = 0;
 531
 532         trans_for_each_update(trans, i) {
 533                 /*
 534                  * XXX: synchronization of cached update triggers with gc
 535                  * XXX: synchronization of interior node updates with gc
 536                  */
 537                 BUG_ON(i->cached || i->level);
 538
 539                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 540                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 541                         if (ret)
 542                                 break;
 543                 }
 544         }
 545
 546         return ret;
 547 }
 548
 549 static inline int
 550 bch2_trans_commit_write_locked(struct btree_trans *trans,
 551                                struct btree_insert_entry **stopped_at,
 552                                unsigned long trace_ip)
 553 {
 554         struct bch_fs *c = trans->c;
 555         struct btree_insert_entry *i;
 556         struct btree_trans_commit_hook *h;
 557         unsigned u64s = 0;
 558         bool marking = false;
 559         int ret;
 560
 561         if (race_fault()) {
 562                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 563                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 564         }
 565
 566         /*
 567          * Check if the insert will fit in the leaf node with the write lock
 568          * held, otherwise another thread could write the node changing the
 569          * amount of space available:
 570          */
 571
 572         prefetch(&trans->c->journal.flags);
 573
 574         h = trans->hooks;
 575         while (h) {
 576                 ret = h->fn(trans, h);
 577                 if (ret)
 578                         return ret;
 579                 h = h->next;
 580         }
 581
 582         trans_for_each_update(trans, i) {
 583                 /* Multiple inserts might go to same leaf: */
 584                 if (!same_leaf_as_prev(trans, i))
 585                         u64s = 0;
 586
 587                 u64s += i->k->k.u64s;
 588                 ret = !i->cached
 589                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 590                         : btree_key_can_insert_cached(trans, i->path, u64s);
 591                 if (ret) {
 592                         *stopped_at = i;
 593                         return ret;
 594                 }
 595
 596                 if (btree_node_type_needs_gc(i->bkey_type))
 597                         marking = true;
 598
 599                 /*
 600                  * Revalidate before calling mem triggers - XXX, ugly:
 601                  *
 602                  * - successful btree node splits don't cause transaction
 603                  *   restarts and will have invalidated the pointer to the bkey
 604                  *   value
 605                  * - btree_node_lock_for_insert() -> btree_node_prep_for_write()
 606                  *   when it has to resort
 607                  * - btree_key_can_insert_cached() when it has to reallocate
 608                  *
 609                  *   Ugly because we currently have no way to tell if the
 610                  *   pointer's been invalidated, which means it's debatabale
 611                  *   whether we should be stashing the old key at all.
 612                  */
 613                 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 614
 615                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
 616                         struct bkey_i *j_k =
 617                                 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 618                                                             i->k->k.p);
 619
 620                         if (j_k) {
 621                                 i->old_k = j_k->k;
 622                                 i->old_v = &j_k->v;
 623                         }
 624                 }
 625         }
 626
 627         /*
 628          * Don't get journal reservation until after we know insert will
 629          * succeed:
 630          */
 631         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 632                 ret = bch2_trans_journal_res_get(trans,
 633                                 JOURNAL_RES_GET_NONBLOCK);
 634                 if (ret)
 635                         return ret;
 636
 637                 journal_transaction_name(trans);
 638         } else {
 639                 trans->journal_res.seq = c->journal.replay_journal_seq;
 640         }
 641
 642         if (unlikely(trans->extra_journal_entries.nr)) {
 643                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 644                                   trans->extra_journal_entries.data,
 645                                   trans->extra_journal_entries.nr);
 646
 647                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 648                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 649         }
 650
 651         /*
 652          * Not allowed to fail after we've gotten our journal reservation - we
 653          * have to use it:
 654          */
 655
 656         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 657                 if (bch2_journal_seq_verify)
 658                         trans_for_each_update(trans, i)
 659                                 i->k->k.version.lo = trans->journal_res.seq;
 660                 else if (bch2_inject_invalid_keys)
 661                         trans_for_each_update(trans, i)
 662                                 i->k->k.version = MAX_VERSION;
 663         }
 664
 665         if (trans->fs_usage_deltas &&
 666             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 667                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 668
 669         trans_for_each_update(trans, i)
 670                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 671                         ret = run_one_mem_trigger(trans, i, i->flags);
 672                         if (ret)
 673                                 return ret;
 674                 }
 675
 676         if (unlikely(c->gc_pos.phase)) {
 677                 ret = bch2_trans_commit_run_gc_triggers(trans);
 678                 if  (ret)
 679                         return ret;
 680         }
 681
 682         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 683                 trans_for_each_update(trans, i) {
 684                         struct journal *j = &c->journal;
 685                         struct jset_entry *entry;
 686
 687                         if (i->key_cache_already_flushed)
 688                                 continue;
 689
 690                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 691                                                BCH_JSET_ENTRY_overwrite,
 692                                                i->btree_id, i->level,
 693                                                i->old_k.u64s);
 694                         bkey_reassemble(&entry->start[0],
 695                                         (struct bkey_s_c) { &i->old_k, i->old_v });
 696
 697                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 698                                                BCH_JSET_ENTRY_btree_keys,
 699                                                i->btree_id, i->level,
 700                                                i->k->k.u64s);
 701                         bkey_copy(&entry->start[0], i->k);
 702                 }
 703
 704                 if (trans->journal_seq)
 705                         *trans->journal_seq = trans->journal_res.seq;
 706         }
 707
 708         trans_for_each_update(trans, i) {
 709                 i->k->k.needs_whiteout = false;
 710
 711                 if (!i->cached)
 712                         btree_insert_key_leaf(trans, i);
 713                 else if (!i->key_cache_already_flushed)
 714                         bch2_btree_insert_key_cached(trans, i->path, i->k);
 715                 else {
 716                         bch2_btree_key_cache_drop(trans, i->path);
 717                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 718                 }
 719         }
 720
 721         return ret;
 722 }
 723
 724 static inline int trans_lock_write(struct btree_trans *trans)
 725 {
 726         struct btree_insert_entry *i;
 727         int ret;
 728
 729         trans_for_each_update(trans, i) {
 730                 if (same_leaf_as_prev(trans, i))
 731                         continue;
 732
 733                 ret = bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c);
 734                 if (ret)
 735                         goto fail;
 736
 737                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 738         }
 739
 740         return 0;
 741 fail:
 742         while (--i >= trans->updates) {
 743                 if (same_leaf_as_prev(trans, i))
 744                         continue;
 745
 746                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 747         }
 748
 749         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 750         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 751 }
 752
 753 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 754 {
 755         struct btree_insert_entry *i;
 756
 757         trans_for_each_update(trans, i)
 758                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 759 }
 760
 761 /*
 762  * Get journal reservation, take write locks, and attempt to do btree update(s):
 763  */
 764 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 765                                        struct btree_insert_entry **stopped_at,
 766                                        unsigned long trace_ip)
 767 {
 768         struct bch_fs *c = trans->c;
 769         struct btree_insert_entry *i;
 770         struct printbuf buf = PRINTBUF;
 771         int ret, u64s_delta = 0;
 772         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 773
 774         trans_for_each_update(trans, i) {
 775                 if (bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 776                                       i->bkey_type, rw, &buf)) {
 777                         printbuf_reset(&buf);
 778                         prt_printf(&buf, "invalid bkey on insert from %s -> %ps",
 779                                trans->fn, (void *) i->ip_allocated);
 780                         prt_newline(&buf);
 781                         printbuf_indent_add(&buf, 2);
 782
 783                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
 784                         prt_newline(&buf);
 785
 786                         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 787                                           i->bkey_type, rw, &buf);
 788
 789                         bch2_trans_inconsistent(trans, "%s", buf.buf);
 790                         printbuf_exit(&buf);
 791                         return -EINVAL;
 792                 }
 793                 btree_insert_entry_checks(trans, i);
 794         }
 795
 796         printbuf_exit(&buf);
 797
 798         trans_for_each_update(trans, i) {
 799                 if (i->cached)
 800                         continue;
 801
 802                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 803                 u64s_delta -= i->old_btree_u64s;
 804
 805                 if (!same_leaf_as_next(trans, i)) {
 806                         if (u64s_delta <= 0) {
 807                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 808                                                         i->level, trans->flags);
 809                                 if (unlikely(ret))
 810                                         return ret;
 811                         }
 812
 813                         u64s_delta = 0;
 814                 }
 815         }
 816
 817         ret = bch2_journal_preres_get(&c->journal,
 818                         &trans->journal_preres, trans->journal_preres_u64s,
 819                         JOURNAL_RES_GET_NONBLOCK|
 820                         (trans->flags & JOURNAL_WATERMARK_MASK));
 821         if (unlikely(ret == -EAGAIN))
 822                 ret = bch2_trans_journal_preres_get_cold(trans,
 823                                                 trans->journal_preres_u64s, trace_ip);
 824         if (unlikely(ret))
 825                 return ret;
 826
 827         ret = trans_lock_write(trans);
 828         if (unlikely(ret))
 829                 return ret;
 830
 831         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 832
 833         if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
 834                 bch2_drop_overwrites_from_journal(trans);
 835
 836         trans_for_each_update(trans, i)
 837                 if (!same_leaf_as_prev(trans, i))
 838                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 839                                                         insert_l(i)->b);
 840
 841         if (!ret && trans->journal_pin)
 842                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 843                                      trans->journal_pin, NULL);
 844
 845         /*
 846          * Drop journal reservation after dropping write locks, since dropping
 847          * the journal reservation may kick off a journal write:
 848          */
 849         bch2_journal_res_put(&c->journal, &trans->journal_res);
 850
 851         if (unlikely(ret))
 852                 return ret;
 853
 854         bch2_trans_downgrade(trans);
 855
 856         return 0;
 857 }
 858
 859 static int journal_reclaim_wait_done(struct bch_fs *c)
 860 {
 861         int ret = bch2_journal_error(&c->journal) ?:
 862                 !bch2_btree_key_cache_must_wait(c);
 863
 864         if (!ret)
 865                 journal_reclaim_kick(&c->journal);
 866         return ret;
 867 }
 868
 869 static noinline
 870 int bch2_trans_commit_error(struct btree_trans *trans,
 871                             struct btree_insert_entry *i,
 872                             int ret, unsigned long trace_ip)
 873 {
 874         struct bch_fs *c = trans->c;
 875
 876         switch (ret) {
 877         case BTREE_INSERT_BTREE_NODE_FULL:
 878                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 879                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 880                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 881                 break;
 882         case BTREE_INSERT_NEED_MARK_REPLICAS:
 883                 bch2_trans_unlock(trans);
 884
 885                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 886                 if (ret)
 887                         break;
 888
 889                 ret = bch2_trans_relock(trans);
 890                 if (ret)
 891                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 892                 break;
 893         case BTREE_INSERT_NEED_JOURNAL_RES:
 894                 bch2_trans_unlock(trans);
 895
 896                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 897                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 898                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 899                         break;
 900                 }
 901
 902                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 903                 if (ret)
 904                         break;
 905
 906                 ret = bch2_trans_relock(trans);
 907                 if (ret)
 908                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 909                 break;
 910         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 911                 bch2_trans_unlock(trans);
 912
 913                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 914
 915                 wait_event_freezable(c->journal.reclaim_wait,
 916                                      (ret = journal_reclaim_wait_done(c)));
 917                 if (ret < 0)
 918                         break;
 919
 920                 ret = bch2_trans_relock(trans);
 921                 if (ret)
 922                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 923                 break;
 924         default:
 925                 BUG_ON(ret >= 0);
 926                 break;
 927         }
 928
 929         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 930
 931         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 932                                 !(trans->flags & BTREE_INSERT_NOWAIT) &&
 933                                 (trans->flags & BTREE_INSERT_NOFAIL), c,
 934                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 935
 936         return ret;
 937 }
 938
 939 static noinline int
 940 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 941 {
 942         struct bch_fs *c = trans->c;
 943         int ret;
 944
 945         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
 946             test_bit(BCH_FS_STARTED, &c->flags))
 947                 return -EROFS;
 948
 949         bch2_trans_unlock(trans);
 950
 951         ret =   bch2_fs_read_write_early(c) ?:
 952                 bch2_trans_relock(trans);
 953         if (ret)
 954                 return ret;
 955
 956         percpu_ref_get(&c->writes);
 957         return 0;
 958 }
 959
 960 /*
 961  * This is for updates done in the early part of fsck - btree_gc - before we've
 962  * gone RW. we only add the new key to the list of keys for journal replay to
 963  * do.
 964  */
 965 static noinline int
 966 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 967 {
 968         struct bch_fs *c = trans->c;
 969         struct btree_insert_entry *i;
 970         int ret = 0;
 971
 972         trans_for_each_update(trans, i) {
 973                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 974                 if (ret)
 975                         break;
 976         }
 977
 978         return ret;
 979 }
 980
 981 int __bch2_trans_commit(struct btree_trans *trans)
 982 {
 983         struct bch_fs *c = trans->c;
 984         struct btree_insert_entry *i = NULL;
 985         unsigned u64s;
 986         int ret = 0;
 987
 988         if (!trans->nr_updates &&
 989             !trans->extra_journal_entries.nr)
 990                 goto out_reset;
 991
 992         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 993                 lockdep_assert_held(&c->gc_lock);
 994
 995         ret = bch2_trans_commit_run_triggers(trans);
 996         if (ret)
 997                 goto out_reset;
 998
 999         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1000                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1001                 goto out_reset;
1002         }
1003
1004         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1005             unlikely(!percpu_ref_tryget_live(&c->writes))) {
1006                 ret = bch2_trans_commit_get_rw_cold(trans);
1007                 if (ret)
1008                         goto out_reset;
1009         }
1010
1011         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1012
1013         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1014
1015         trans->journal_u64s             = trans->extra_journal_entries.nr;
1016         trans->journal_preres_u64s      = 0;
1017
1018         /* For journalling transaction name: */
1019         trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1020
1021         trans_for_each_update(trans, i) {
1022                 BUG_ON(!i->path->should_be_locked);
1023
1024                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1025                 if (unlikely(ret))
1026                         goto out;
1027
1028                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1029
1030                 if (i->key_cache_already_flushed)
1031                         continue;
1032
1033                 /* we're going to journal the key being updated: */
1034                 u64s = jset_u64s(i->k->k.u64s);
1035                 if (i->cached &&
1036                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1037                         trans->journal_preres_u64s += u64s;
1038                 trans->journal_u64s += u64s;
1039
1040                 /* and we're also going to log the overwrite: */
1041                 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1042         }
1043
1044         if (trans->extra_journal_res) {
1045                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1046                                 trans->extra_journal_res,
1047                                 (trans->flags & BTREE_INSERT_NOFAIL)
1048                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1049                 if (ret)
1050                         goto err;
1051         }
1052 retry:
1053         BUG_ON(trans->restarted);
1054         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1055
1056         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1057
1058         /* make sure we didn't drop or screw up locks: */
1059         bch2_trans_verify_locks(trans);
1060
1061         if (ret)
1062                 goto err;
1063
1064         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1065 out:
1066         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1067
1068         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1069                 percpu_ref_put(&c->writes);
1070 out_reset:
1071         bch2_trans_reset_updates(trans);
1072
1073         if (trans->fs_usage_deltas) {
1074                 trans->fs_usage_deltas->used = 0;
1075                 memset((void *) trans->fs_usage_deltas +
1076                        offsetof(struct replicas_delta_list, memset_start), 0,
1077                        (void *) &trans->fs_usage_deltas->memset_end -
1078                        (void *) &trans->fs_usage_deltas->memset_start);
1079         }
1080
1081         return ret;
1082 err:
1083         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1084         if (ret)
1085                 goto out;
1086
1087         goto retry;
1088 }
1089
1090 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1091                                           enum btree_id id,
1092                                           struct bpos pos)
1093 {
1094         struct bch_fs *c = trans->c;
1095         struct btree_iter iter;
1096         struct bkey_s_c k;
1097         int ret;
1098
1099         bch2_trans_iter_init(trans, &iter, id, pos,
1100                              BTREE_ITER_NOT_EXTENTS|
1101                              BTREE_ITER_ALL_SNAPSHOTS);
1102         while (1) {
1103                 k = bch2_btree_iter_prev(&iter);
1104                 ret = bkey_err(k);
1105                 if (ret)
1106                         break;
1107
1108                 if (!k.k)
1109                         break;
1110
1111                 if (bkey_cmp(pos, k.k->p))
1112                         break;
1113
1114                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1115                         ret = 1;
1116                         break;
1117                 }
1118         }
1119         bch2_trans_iter_exit(trans, &iter);
1120
1121         return ret;
1122 }
1123
1124 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1125                                           enum btree_id id,
1126                                           struct bpos pos)
1127 {
1128         if (!btree_type_has_snapshots(id) ||
1129             pos.snapshot == U32_MAX ||
1130             !snapshot_t(trans->c, pos.snapshot)->children[0])
1131                 return 0;
1132
1133         return __check_pos_snapshot_overwritten(trans, id, pos);
1134 }
1135
1136 int bch2_trans_update_extent(struct btree_trans *trans,
1137                              struct btree_iter *orig_iter,
1138                              struct bkey_i *insert,
1139                              enum btree_update_flags flags)
1140 {
1141         struct bch_fs *c = trans->c;
1142         struct btree_iter iter, update_iter;
1143         struct bpos start = bkey_start_pos(&insert->k);
1144         struct bkey_i *update;
1145         struct bkey_s_c k;
1146         enum btree_id btree_id = orig_iter->btree_id;
1147         int ret = 0, compressed_sectors;
1148
1149         bch2_trans_iter_init(trans, &iter, btree_id, start,
1150                              BTREE_ITER_INTENT|
1151                              BTREE_ITER_WITH_UPDATES|
1152                              BTREE_ITER_NOT_EXTENTS);
1153         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1154         if ((ret = bkey_err(k)))
1155                 goto err;
1156         if (!k.k)
1157                 goto out;
1158
1159         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1160                 /*
1161                  * We can't merge extents if they belong to interior snapshot
1162                  * tree nodes, and there's a snapshot in which one extent is
1163                  * visible and the other is not - i.e. if visibility is
1164                  * different.
1165                  *
1166                  * Instead of checking if visibilitiy of the two extents is
1167                  * different, for now we just check if either has been
1168                  * overwritten:
1169                  */
1170                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1171                 if (ret < 0)
1172                         goto err;
1173                 if (ret)
1174                         goto nomerge1;
1175
1176                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1177                 if (ret < 0)
1178                         goto err;
1179                 if (ret)
1180                         goto nomerge1;
1181
1182                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1183                 if ((ret = PTR_ERR_OR_ZERO(update)))
1184                         goto err;
1185
1186                 bkey_reassemble(update, k);
1187
1188                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1189                         ret = bch2_btree_delete_at(trans, &iter, flags);
1190                         if (ret)
1191                                 goto err;
1192
1193                         insert = update;
1194                         goto next;
1195                 }
1196         }
1197 nomerge1:
1198         ret = 0;
1199         if (!bkey_cmp(k.k->p, start))
1200                 goto next;
1201
1202         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1203                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1204                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1205
1206                 /*
1207                  * If we're going to be splitting a compressed extent, note it
1208                  * so that __bch2_trans_commit() can increase our disk
1209                  * reservation:
1210                  */
1211                 if (((front_split && back_split) ||
1212                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1213                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1214                         trans->extra_journal_res += compressed_sectors;
1215
1216                 if (front_split) {
1217                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1218                         if ((ret = PTR_ERR_OR_ZERO(update)))
1219                                 goto err;
1220
1221                         bkey_reassemble(update, k);
1222
1223                         bch2_cut_back(start, update);
1224
1225                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1226                                              BTREE_ITER_NOT_EXTENTS|
1227                                              BTREE_ITER_ALL_SNAPSHOTS|
1228                                              BTREE_ITER_INTENT);
1229                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1230                                 bch2_trans_update(trans, &update_iter, update,
1231                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1232                                                   flags);
1233                         bch2_trans_iter_exit(trans, &update_iter);
1234
1235                         if (ret)
1236                                 goto err;
1237                 }
1238
1239                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1240                     (front_split || back_split)) {
1241                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1242                         if ((ret = PTR_ERR_OR_ZERO(update)))
1243                                 goto err;
1244
1245                         bkey_reassemble(update, k);
1246
1247                         bch2_cut_front(start, update);
1248                         bch2_cut_back(insert->k.p, update);
1249
1250                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1251                                              BTREE_ITER_NOT_EXTENTS|
1252                                              BTREE_ITER_ALL_SNAPSHOTS|
1253                                              BTREE_ITER_INTENT);
1254                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1255                                 bch2_trans_update(trans, &update_iter, update,
1256                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1257                                                   flags);
1258                         bch2_trans_iter_exit(trans, &update_iter);
1259                         if (ret)
1260                                 goto err;
1261                 }
1262
1263                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1264                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1265                         if ((ret = PTR_ERR_OR_ZERO(update)))
1266                                 goto err;
1267
1268                         bkey_init(&update->k);
1269                         update->k.p = k.k->p;
1270
1271                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1272                                 update->k.p.snapshot = insert->k.p.snapshot;
1273                                 update->k.type = KEY_TYPE_whiteout;
1274                         }
1275
1276                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1277                                              BTREE_ITER_NOT_EXTENTS|
1278                                              BTREE_ITER_INTENT);
1279                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1280                                 bch2_trans_update(trans, &update_iter, update,
1281                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1282                                                   flags);
1283                         bch2_trans_iter_exit(trans, &update_iter);
1284
1285                         if (ret)
1286                                 goto err;
1287                 }
1288
1289                 if (back_split) {
1290                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1291                         if ((ret = PTR_ERR_OR_ZERO(update)))
1292                                 goto err;
1293
1294                         bkey_reassemble(update, k);
1295                         bch2_cut_front(insert->k.p, update);
1296
1297                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1298                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1299                                                   flags);
1300                         if (ret)
1301                                 goto err;
1302                         goto out;
1303                 }
1304 next:
1305                 bch2_btree_iter_advance(&iter);
1306                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1307                 if ((ret = bkey_err(k)))
1308                         goto err;
1309                 if (!k.k)
1310                         goto out;
1311         }
1312
1313         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1314                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1315                 if (ret < 0)
1316                         goto err;
1317                 if (ret)
1318                         goto nomerge2;
1319
1320                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1321                 if (ret < 0)
1322                         goto err;
1323                 if (ret)
1324                         goto nomerge2;
1325
1326                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1327         }
1328 nomerge2:
1329         ret = 0;
1330 out:
1331         if (!bkey_deleted(&insert->k)) {
1332                 /*
1333                  * Rewinding iterators is expensive: get a new one and the one
1334                  * that points to the start of insert will be cloned from:
1335                  */
1336                 bch2_trans_iter_exit(trans, &iter);
1337                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1338                                      BTREE_ITER_NOT_EXTENTS|
1339                                      BTREE_ITER_INTENT);
1340                 ret   = bch2_btree_iter_traverse(&iter) ?:
1341                         bch2_trans_update(trans, &iter, insert, flags);
1342         }
1343 err:
1344         bch2_trans_iter_exit(trans, &iter);
1345
1346         return ret;
1347 }
1348
1349 /*
1350  * When deleting, check if we need to emit a whiteout (because we're overwriting
1351  * something in an ancestor snapshot)
1352  */
1353 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1354                                       enum btree_id btree_id, struct bpos pos)
1355 {
1356         struct btree_iter iter;
1357         struct bkey_s_c k;
1358         u32 snapshot = pos.snapshot;
1359         int ret;
1360
1361         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1362                 return 0;
1363
1364         pos.snapshot++;
1365
1366         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1367                            BTREE_ITER_ALL_SNAPSHOTS|
1368                            BTREE_ITER_NOPRESERVE, k, ret) {
1369                 if (bkey_cmp(k.k->p, pos))
1370                         break;
1371
1372                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1373                                               k.k->p.snapshot)) {
1374                         ret = !bkey_whiteout(k.k);
1375                         break;
1376                 }
1377         }
1378         bch2_trans_iter_exit(trans, &iter);
1379
1380         return ret;
1381 }
1382
1383 static int __must_check
1384 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1385                                 struct bkey_i *k, enum btree_update_flags flags,
1386                                 unsigned long ip);
1387
1388 static noinline int flush_new_cached_update(struct btree_trans *trans,
1389                                             struct btree_path *path,
1390                                             struct btree_insert_entry *i,
1391                                             enum btree_update_flags flags,
1392                                             unsigned long ip)
1393 {
1394         struct btree_path *btree_path;
1395         int ret;
1396
1397         i->key_cache_already_flushed = true;
1398         i->flags |= BTREE_TRIGGER_NORUN;
1399
1400         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1401                                    BTREE_ITER_INTENT, _THIS_IP_);
1402
1403         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1404         if (ret)
1405                 goto err;
1406
1407         btree_path_set_should_be_locked(btree_path);
1408         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1409 err:
1410         bch2_path_put(trans, btree_path, true);
1411         return ret;
1412 }
1413
1414 static int __must_check
1415 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1416                                 struct bkey_i *k, enum btree_update_flags flags,
1417                                 unsigned long ip)
1418 {
1419         struct bch_fs *c = trans->c;
1420         struct btree_insert_entry *i, n;
1421
1422         BUG_ON(!path->should_be_locked);
1423
1424         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1425         BUG_ON(bpos_cmp(k->k.p, path->pos));
1426
1427         n = (struct btree_insert_entry) {
1428                 .flags          = flags,
1429                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1430                 .btree_id       = path->btree_id,
1431                 .level          = path->level,
1432                 .cached         = path->cached,
1433                 .path           = path,
1434                 .k              = k,
1435                 .ip_allocated   = ip,
1436         };
1437
1438 #ifdef CONFIG_BCACHEFS_DEBUG
1439         trans_for_each_update(trans, i)
1440                 BUG_ON(i != trans->updates &&
1441                        btree_insert_entry_cmp(i - 1, i) >= 0);
1442 #endif
1443
1444         /*
1445          * Pending updates are kept sorted: first, find position of new update,
1446          * then delete/trim any updates the new update overwrites:
1447          */
1448         trans_for_each_update(trans, i)
1449                 if (btree_insert_entry_cmp(&n, i) <= 0)
1450                         break;
1451
1452         if (i < trans->updates + trans->nr_updates &&
1453             !btree_insert_entry_cmp(&n, i)) {
1454                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1455
1456                 bch2_path_put(trans, i->path, true);
1457                 i->flags        = n.flags;
1458                 i->cached       = n.cached;
1459                 i->k            = n.k;
1460                 i->path         = n.path;
1461                 i->ip_allocated = n.ip_allocated;
1462         } else {
1463                 array_insert_item(trans->updates, trans->nr_updates,
1464                                   i - trans->updates, n);
1465
1466                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1467                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1468
1469                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
1470                         struct bkey_i *j_k =
1471                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1472
1473                         if (j_k) {
1474                                 i->old_k = j_k->k;
1475                                 i->old_v = &j_k->v;
1476                         }
1477                 }
1478         }
1479
1480         __btree_path_get(i->path, true);
1481
1482         /*
1483          * If a key is present in the key cache, it must also exist in the
1484          * btree - this is necessary for cache coherency. When iterating over
1485          * a btree that's cached in the key cache, the btree iter code checks
1486          * the key cache - but the key has to exist in the btree for that to
1487          * work:
1488          */
1489         if (unlikely(path->cached && bkey_deleted(&i->old_k)))
1490                 return flush_new_cached_update(trans, path, i, flags, ip);
1491
1492         return 0;
1493 }
1494
1495 static int __must_check
1496 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1497                           struct bkey_i *k, enum btree_update_flags flags)
1498 {
1499         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1500 }
1501
1502 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1503                                    struct bkey_i *k, enum btree_update_flags flags)
1504 {
1505         struct btree_path *path = iter->update_path ?: iter->path;
1506         struct bkey_cached *ck;
1507         int ret;
1508
1509         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1510                 return bch2_trans_update_extent(trans, iter, k, flags);
1511
1512         if (bkey_deleted(&k->k) &&
1513             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1514             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1515                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1516                 if (unlikely(ret < 0))
1517                         return ret;
1518
1519                 if (ret)
1520                         k->k.type = KEY_TYPE_whiteout;
1521         }
1522
1523         /*
1524          * Ensure that updates to cached btrees go to the key cache:
1525          */
1526         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1527             !path->cached &&
1528             !path->level &&
1529             btree_id_cached(trans->c, path->btree_id)) {
1530                 if (!iter->key_cache_path ||
1531                     !iter->key_cache_path->should_be_locked ||
1532                     bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
1533                         if (!iter->key_cache_path)
1534                                 iter->key_cache_path =
1535                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1536                                                       BTREE_ITER_INTENT|
1537                                                       BTREE_ITER_CACHED, _THIS_IP_);
1538
1539                         iter->key_cache_path =
1540                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1541                                                         iter->flags & BTREE_ITER_INTENT,
1542                                                         _THIS_IP_);
1543
1544                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1545                                                        BTREE_ITER_CACHED);
1546                         if (unlikely(ret))
1547                                 return ret;
1548
1549                         ck = (void *) iter->key_cache_path->l[0].b;
1550
1551                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1552                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1553                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1554                         }
1555
1556                         btree_path_set_should_be_locked(iter->key_cache_path);
1557                 }
1558
1559                 path = iter->key_cache_path;
1560         }
1561
1562         return bch2_trans_update_by_path(trans, path, k, flags);
1563 }
1564
1565 void bch2_trans_commit_hook(struct btree_trans *trans,
1566                             struct btree_trans_commit_hook *h)
1567 {
1568         h->next = trans->hooks;
1569         trans->hooks = h;
1570 }
1571
1572 int __bch2_btree_insert(struct btree_trans *trans,
1573                         enum btree_id id, struct bkey_i *k)
1574 {
1575         struct btree_iter iter;
1576         int ret;
1577
1578         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1579                              BTREE_ITER_INTENT);
1580         ret   = bch2_btree_iter_traverse(&iter) ?:
1581                 bch2_trans_update(trans, &iter, k, 0);
1582         bch2_trans_iter_exit(trans, &iter);
1583         return ret;
1584 }
1585
1586 /**
1587  * bch2_btree_insert - insert keys into the extent btree
1588  * @c:                  pointer to struct bch_fs
1589  * @id:                 btree to insert into
1590  * @insert_keys:        list of keys to insert
1591  * @hook:               insert callback
1592  */
1593 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1594                       struct bkey_i *k,
1595                       struct disk_reservation *disk_res,
1596                       u64 *journal_seq, int flags)
1597 {
1598         return bch2_trans_do(c, disk_res, journal_seq, flags,
1599                              __bch2_btree_insert(&trans, id, k));
1600 }
1601
1602 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1603                                 unsigned len, unsigned update_flags)
1604 {
1605         struct bkey_i *k;
1606
1607         k = bch2_trans_kmalloc(trans, sizeof(*k));
1608         if (IS_ERR(k))
1609                 return PTR_ERR(k);
1610
1611         bkey_init(&k->k);
1612         k->k.p = iter->pos;
1613         bch2_key_resize(&k->k, len);
1614         return bch2_trans_update(trans, iter, k, update_flags);
1615 }
1616
1617 int bch2_btree_delete_at(struct btree_trans *trans,
1618                          struct btree_iter *iter, unsigned update_flags)
1619 {
1620         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1621 }
1622
1623 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1624                                   struct bpos start, struct bpos end,
1625                                   unsigned update_flags,
1626                                   u64 *journal_seq)
1627 {
1628         u32 restart_count = trans->restart_count;
1629         struct btree_iter iter;
1630         struct bkey_s_c k;
1631         int ret = 0;
1632
1633         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1634         while ((k = bch2_btree_iter_peek(&iter)).k) {
1635                 struct disk_reservation disk_res =
1636                         bch2_disk_reservation_init(trans->c, 0);
1637                 struct bkey_i delete;
1638
1639                 ret = bkey_err(k);
1640                 if (ret)
1641                         goto err;
1642
1643                 if (bkey_cmp(iter.pos, end) >= 0)
1644                         break;
1645
1646                 bkey_init(&delete.k);
1647
1648                 /*
1649                  * This could probably be more efficient for extents:
1650                  */
1651
1652                 /*
1653                  * For extents, iter.pos won't necessarily be the same as
1654                  * bkey_start_pos(k.k) (for non extents they always will be the
1655                  * same). It's important that we delete starting from iter.pos
1656                  * because the range we want to delete could start in the middle
1657                  * of k.
1658                  *
1659                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1660                  * bkey_start_pos(k.k)).
1661                  */
1662                 delete.k.p = iter.pos;
1663
1664                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1665                         unsigned max_sectors =
1666                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1667
1668                         /* create the biggest key we can */
1669                         bch2_key_resize(&delete.k, max_sectors);
1670                         bch2_cut_back(end, &delete);
1671
1672                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1673                         if (ret)
1674                                 goto err;
1675                 }
1676
1677                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1678                         bch2_trans_commit(trans, &disk_res, journal_seq,
1679                                           BTREE_INSERT_NOFAIL);
1680                 bch2_disk_reservation_put(trans->c, &disk_res);
1681 err:
1682                 /*
1683                  * the bch2_trans_begin() call is in a weird place because we
1684                  * need to call it after every transaction commit, to avoid path
1685                  * overflow, but don't want to call it if the delete operation
1686                  * is a no-op and we have no work to do:
1687                  */
1688                 bch2_trans_begin(trans);
1689
1690                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1691                         ret = 0;
1692                 if (ret)
1693                         break;
1694         }
1695         bch2_trans_iter_exit(trans, &iter);
1696
1697         if (!ret && trans_was_restarted(trans, restart_count))
1698                 ret = -BCH_ERR_transaction_restart_nested;
1699         return ret;
1700 }
1701
1702 /*
1703  * bch_btree_delete_range - delete everything within a given range
1704  *
1705  * Range is a half open interval - [start, end)
1706  */
1707 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1708                             struct bpos start, struct bpos end,
1709                             unsigned update_flags,
1710                             u64 *journal_seq)
1711 {
1712         int ret = bch2_trans_run(c,
1713                         bch2_btree_delete_range_trans(&trans, id, start, end,
1714                                                       update_flags, journal_seq));
1715         if (ret == -BCH_ERR_transaction_restart_nested)
1716                 ret = 0;
1717         return ret;
1718 }
1719
1720 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1721 {
1722         unsigned len = strlen(msg);
1723         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1724         struct jset_entry_log *l;
1725         int ret;
1726
1727         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1728         if (ret)
1729                 return ret;
1730
1731         l = (void *) &darray_top(trans->extra_journal_entries);
1732         l->entry.u64s           = cpu_to_le16(u64s);
1733         l->entry.btree_id       = 0;
1734         l->entry.level          = 1;
1735         l->entry.type           = BCH_JSET_ENTRY_log;
1736         l->entry.pad[0]         = 0;
1737         l->entry.pad[1]         = 0;
1738         l->entry.pad[2]         = 0;
1739         memcpy(l->d, msg, len);
1740         while (len & 7)
1741                 l->d[len++] = '\0';
1742
1743         trans->extra_journal_entries.nr += jset_u64s(u64s);
1744         return 0;
1745 }