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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_iter.h"
   6 #include "btree_journal_iter.h"
   7 #include "btree_locking.h"
   8 #include "buckets.h"
   9 #include "debug.h"
  10 #include "errcode.h"
  11 #include "error.h"
  12 #include "extents.h"
  13 #include "keylist.h"
  14 #include "snapshot.h"
  15 #include "trace.h"
  16
  17 #include <linux/darray.h>
  18
  19 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  20                                          const struct btree_insert_entry *r)
  21 {
  22         return   cmp_int(l->btree_id,   r->btree_id) ?:
  23                  cmp_int(l->cached,     r->cached) ?:
  24                  -cmp_int(l->level,     r->level) ?:
  25                  bpos_cmp(l->k->k.p,    r->k->k.p);
  26 }
  27
  28 static int __must_check
  29 bch2_trans_update_by_path(struct btree_trans *, btree_path_idx_t,
  30                           struct bkey_i *, enum btree_update_flags,
  31                           unsigned long ip);
  32
  33 static noinline int extent_front_merge(struct btree_trans *trans,
  34                                        struct btree_iter *iter,
  35                                        struct bkey_s_c k,
  36                                        struct bkey_i **insert,
  37                                        enum btree_update_flags flags)
  38 {
  39         struct bch_fs *c = trans->c;
  40         struct bkey_i *update;
  41         int ret;
  42
  43         update = bch2_bkey_make_mut_noupdate(trans, k);
  44         ret = PTR_ERR_OR_ZERO(update);
  45         if (ret)
  46                 return ret;
  47
  48         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
  49                 return 0;
  50
  51         ret =   bch2_key_has_snapshot_overwrites(trans, iter->btree_id, k.k->p) ?:
  52                 bch2_key_has_snapshot_overwrites(trans, iter->btree_id, (*insert)->k.p);
  53         if (ret < 0)
  54                 return ret;
  55         if (ret)
  56                 return 0;
  57
  58         ret = bch2_btree_delete_at(trans, iter, flags);
  59         if (ret)
  60                 return ret;
  61
  62         *insert = update;
  63         return 0;
  64 }
  65
  66 static noinline int extent_back_merge(struct btree_trans *trans,
  67                                       struct btree_iter *iter,
  68                                       struct bkey_i *insert,
  69                                       struct bkey_s_c k)
  70 {
  71         struct bch_fs *c = trans->c;
  72         int ret;
  73
  74         ret =   bch2_key_has_snapshot_overwrites(trans, iter->btree_id, insert->k.p) ?:
  75                 bch2_key_has_snapshot_overwrites(trans, iter->btree_id, k.k->p);
  76         if (ret < 0)
  77                 return ret;
  78         if (ret)
  79                 return 0;
  80
  81         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
  82         return 0;
  83 }
  84
  85 static struct bkey_s_c peek_slot_including_whiteouts(struct btree_trans *trans, struct btree_iter *iter,
  86                                                      enum btree_id btree, struct bpos pos)
  87 {
  88         struct bkey_s_c k;
  89         int ret;
  90
  91         for_each_btree_key_norestart(trans, *iter, btree, pos,
  92                            BTREE_ITER_ALL_SNAPSHOTS|
  93                            BTREE_ITER_NOPRESERVE, k, ret) {
  94                 if (!bkey_eq(k.k->p, pos))
  95                         break;
  96                 if (bch2_snapshot_is_ancestor(trans->c, pos.snapshot, k.k->p.snapshot))
  97                         return k;
  98         }
  99         bch2_trans_iter_exit(trans, iter);
 100
 101         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 102 }
 103
 104 /*
 105  * When deleting, check if we need to emit a whiteout (because we're overwriting
 106  * something in an ancestor snapshot)
 107  */
 108 static int need_whiteout_for_snapshot(struct btree_trans *trans, enum btree_id btree, struct bpos pos)
 109 {
 110         pos.snapshot = bch2_snapshot_parent(trans->c, pos.snapshot);
 111         if (!pos.snapshot)
 112                 return 0;
 113
 114         struct btree_iter iter;
 115         struct bkey_s_c k = peek_slot_including_whiteouts(trans, &iter, btree, pos);
 116         int ret = bkey_err(k) ?: k.k && !bkey_whiteout(k.k);
 117         bch2_trans_iter_exit(trans, &iter);
 118
 119         return ret;
 120 }
 121
 122 /*
 123  * We're overwriting a key at @pos in snapshot @snapshot, so we need to insert a
 124  * whiteout: that might be in @snapshot, or if there are overwites in sibling
 125  * snapshots, find the common ancestor where @pos is overwritten in every
 126  * descendent and insert the whiteout there - which might be at @pos.
 127  */
 128 static int delete_interior_snapshot_key(struct btree_trans *trans,
 129                                         enum btree_id btree,
 130                                         struct bpos whiteout, bool deleting,
 131                                         struct bpos overwrite, bool old_is_whiteout)
 132 {
 133         struct bch_fs *c = trans->c;
 134         struct bpos orig_whiteout = whiteout, sib = whiteout;
 135         struct btree_iter iter;
 136         struct bkey_s_c k;
 137         int ret;
 138
 139         sib.snapshot = bch2_snapshot_sibling(c, sib.snapshot);
 140
 141         for_each_btree_key_norestart(trans, iter, btree, sib,
 142                                      BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_INTENT, k, ret) {
 143                 BUG_ON(bpos_gt(k.k->p, overwrite));
 144
 145                 if (bpos_lt(k.k->p, sib)) /* unrelated branch - skip */
 146                         continue;
 147                 if (bpos_gt(k.k->p, sib)) /* did not find @sib */
 148                         break;
 149
 150                 /* @overwrite is also written in @sib, now check parent */
 151                 whiteout.snapshot = bch2_snapshot_parent(c, whiteout.snapshot);
 152                 if (bpos_eq(whiteout, overwrite))
 153                         break;
 154
 155                 sib = whiteout;
 156                 sib.snapshot = bch2_snapshot_sibling(c, sib.snapshot);
 157         }
 158
 159         if (ret)
 160                 goto err;
 161
 162         if (!deleting && bpos_eq(whiteout, orig_whiteout))
 163                 goto out;
 164
 165         if (!bpos_eq(iter.pos, whiteout)) {
 166                 bch2_trans_iter_exit(trans, &iter);
 167                 bch2_trans_iter_init(trans, &iter, btree, whiteout, BTREE_ITER_INTENT);
 168                 k = bch2_btree_iter_peek_slot(&iter);
 169                 ret = bkey_err(k);
 170                 if (ret)
 171                         goto err;
 172         }
 173
 174         iter.flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
 175         iter.flags |= BTREE_ITER_FILTER_SNAPSHOTS;
 176
 177         struct bkey_i *delete = bch2_trans_kmalloc(trans, sizeof(*delete));
 178         ret = PTR_ERR_OR_ZERO(delete);
 179         if (ret)
 180                 goto err;
 181
 182         bkey_init(&delete->k);
 183         delete->k.p = whiteout;
 184
 185         ret = !bpos_eq(whiteout, overwrite)
 186                 ? !old_is_whiteout
 187                 : need_whiteout_for_snapshot(trans, btree, whiteout);
 188         if (ret < 0)
 189                 goto err;
 190         if (ret)
 191                 delete->k.type = KEY_TYPE_whiteout;
 192
 193         ret = bch2_trans_update(trans, &iter, delete,
 194                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 195                                 BTREE_UPDATE_SNAPSHOT_WHITEOUT_CHECKS_DONE);
 196 out:
 197 err:
 198         bch2_trans_iter_exit(trans, &iter);
 199         return ret;
 200 }
 201
 202 /*
 203  * We're overwriting a key in a snapshot that has ancestors: if we're
 204  * overwriting a key in a different snapshot, we need to check if it is now
 205  * fully overritten and can be deleted, and if we're deleting a key in the
 206  * current snapshot we need to check if we need to leave a whiteout.
 207  */
 208 static noinline int
 209 overwrite_interior_snapshot_key(struct btree_trans *trans,
 210                                 struct btree_iter *iter,
 211                                 struct bkey_i *k)
 212 {
 213         struct bkey_s_c old = bch2_btree_iter_peek_slot(iter);
 214
 215         int ret = bkey_err(old);
 216         if (ret)
 217                 return ret;
 218
 219         if (!bkey_deleted(old.k)) {
 220                 if (old.k->p.snapshot != k->k.p.snapshot) {
 221                         /*
 222                          * We're overwriting a key in a different snapshot:
 223                          * check if it's also been overwritten in siblings
 224                          */
 225                         ret = delete_interior_snapshot_key(trans, iter->btree_id,
 226                                                            k->k.p,   bkey_deleted(&k->k),
 227                                                            old.k->p, bkey_whiteout(old.k));
 228                         if (ret)
 229                                 return ret;
 230                         if (bkey_deleted(&k->k))
 231                                 return 1;
 232                 } else if (bkey_deleted(&k->k)) {
 233                         /*
 234                          * We're deleting a key in the current snapshot:
 235                          * check if we need to leave a whiteout
 236                          */
 237                         ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
 238                         if (unlikely(ret < 0))
 239                                 return ret;
 240                         if (ret)
 241                                 k->k.type = KEY_TYPE_whiteout;
 242                 }
 243         }
 244
 245         return 0;
 246 }
 247
 248 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
 249                                    enum btree_id id,
 250                                    struct bpos old_pos,
 251                                    struct bpos new_pos)
 252 {
 253         struct bch_fs *c = trans->c;
 254         struct btree_iter old_iter, new_iter = { NULL };
 255         struct bkey_s_c old_k, new_k;
 256         snapshot_id_list s;
 257         struct bkey_i *update;
 258         int ret = 0;
 259
 260         if (!bch2_snapshot_has_children(c, old_pos.snapshot))
 261                 return 0;
 262
 263         darray_init(&s);
 264
 265         bch2_trans_iter_init(trans, &old_iter, id, old_pos,
 266                              BTREE_ITER_NOT_EXTENTS|
 267                              BTREE_ITER_ALL_SNAPSHOTS);
 268         while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
 269                !(ret = bkey_err(old_k)) &&
 270                bkey_eq(old_pos, old_k.k->p)) {
 271                 struct bpos whiteout_pos =
 272                         SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
 273
 274                 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
 275                     snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
 276                         continue;
 277
 278                 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
 279                                            BTREE_ITER_NOT_EXTENTS|
 280                                            BTREE_ITER_INTENT);
 281                 ret = bkey_err(new_k);
 282                 if (ret)
 283                         break;
 284
 285                 if (new_k.k->type == KEY_TYPE_deleted) {
 286                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
 287                         ret = PTR_ERR_OR_ZERO(update);
 288                         if (ret)
 289                                 break;
 290
 291                         bkey_init(&update->k);
 292                         update->k.p             = whiteout_pos;
 293                         update->k.type          = KEY_TYPE_whiteout;
 294
 295                         ret = bch2_trans_update(trans, &new_iter, update,
 296                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 297                 }
 298                 bch2_trans_iter_exit(trans, &new_iter);
 299
 300                 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
 301                 if (ret)
 302                         break;
 303         }
 304         bch2_trans_iter_exit(trans, &new_iter);
 305         bch2_trans_iter_exit(trans, &old_iter);
 306         darray_exit(&s);
 307
 308         return ret;
 309 }
 310
 311 int bch2_trans_update_extent_overwrite(struct btree_trans *trans,
 312                                        struct btree_iter *iter,
 313                                        enum btree_update_flags flags,
 314                                        struct bkey_s_c old,
 315                                        struct bkey_s_c new)
 316 {
 317         enum btree_id btree_id = iter->btree_id;
 318         struct bkey_i *update;
 319         struct bpos new_start = bkey_start_pos(new.k);
 320         unsigned front_split = bkey_lt(bkey_start_pos(old.k), new_start);
 321         unsigned back_split  = bkey_gt(old.k->p, new.k->p);
 322         unsigned middle_split = (front_split || back_split) &&
 323                 old.k->p.snapshot != new.k->p.snapshot;
 324         unsigned nr_splits = front_split + back_split + middle_split;
 325         int ret = 0, compressed_sectors;
 326
 327         /*
 328          * If we're going to be splitting a compressed extent, note it
 329          * so that __bch2_trans_commit() can increase our disk
 330          * reservation:
 331          */
 332         if (nr_splits > 1 &&
 333             (compressed_sectors = bch2_bkey_sectors_compressed(old)))
 334                 trans->extra_disk_res += compressed_sectors * (nr_splits - 1);
 335
 336         if (front_split) {
 337                 update = bch2_bkey_make_mut_noupdate(trans, old);
 338                 if ((ret = PTR_ERR_OR_ZERO(update)))
 339                         return ret;
 340
 341                 bch2_cut_back(new_start, update);
 342
 343                 ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
 344                                         old.k->p, update->k.p) ?:
 345                         bch2_btree_insert_nonextent(trans, btree_id, update,
 346                                         BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 347                 if (ret)
 348                         return ret;
 349         }
 350
 351         /* If we're overwriting in a different snapshot - middle split: */
 352         if (middle_split) {
 353                 update = bch2_bkey_make_mut_noupdate(trans, old);
 354                 if ((ret = PTR_ERR_OR_ZERO(update)))
 355                         return ret;
 356
 357                 bch2_cut_front(new_start, update);
 358                 bch2_cut_back(new.k->p, update);
 359
 360                 ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
 361                                         old.k->p, update->k.p) ?:
 362                         bch2_btree_insert_nonextent(trans, btree_id, update,
 363                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 364                 if (ret)
 365                         return ret;
 366         }
 367
 368         if (bkey_le(old.k->p, new.k->p)) {
 369                 update = bch2_trans_kmalloc(trans, sizeof(*update));
 370                 if ((ret = PTR_ERR_OR_ZERO(update)))
 371                         return ret;
 372
 373                 bkey_init(&update->k);
 374                 update->k.p = old.k->p;
 375                 update->k.p.snapshot = new.k->p.snapshot;
 376
 377                 if (new.k->p.snapshot != old.k->p.snapshot) {
 378                         update->k.type = KEY_TYPE_whiteout;
 379                 } else if (btree_type_has_snapshots(btree_id)) {
 380                         ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
 381                         if (ret < 0)
 382                                 return ret;
 383                         if (ret)
 384                                 update->k.type = KEY_TYPE_whiteout;
 385                 }
 386
 387                 ret = bch2_btree_insert_nonextent(trans, btree_id, update,
 388                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 389                 if (ret)
 390                         return ret;
 391         }
 392
 393         if (back_split) {
 394                 update = bch2_bkey_make_mut_noupdate(trans, old);
 395                 if ((ret = PTR_ERR_OR_ZERO(update)))
 396                         return ret;
 397
 398                 bch2_cut_front(new.k->p, update);
 399
 400                 ret = bch2_trans_update_by_path(trans, iter->path, update,
 401                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 402                                           flags, _RET_IP_);
 403                 if (ret)
 404                         return ret;
 405         }
 406
 407         return 0;
 408 }
 409
 410 static int bch2_trans_update_extent(struct btree_trans *trans,
 411                                     struct btree_iter *orig_iter,
 412                                     struct bkey_i *insert,
 413                                     enum btree_update_flags flags)
 414 {
 415         struct btree_iter iter;
 416         struct bkey_s_c k;
 417         enum btree_id btree_id = orig_iter->btree_id;
 418         int ret = 0;
 419
 420         bch2_trans_iter_init(trans, &iter, btree_id, bkey_start_pos(&insert->k),
 421                              BTREE_ITER_INTENT|
 422                              BTREE_ITER_WITH_UPDATES|
 423                              BTREE_ITER_NOT_EXTENTS);
 424         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
 425         if ((ret = bkey_err(k)))
 426                 goto err;
 427         if (!k.k)
 428                 goto out;
 429
 430         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
 431                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
 432                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
 433                         if (ret)
 434                                 goto err;
 435                 }
 436
 437                 goto next;
 438         }
 439
 440         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
 441                 bool done = bkey_lt(insert->k.p, k.k->p);
 442
 443                 ret = bch2_trans_update_extent_overwrite(trans, &iter, flags, k, bkey_i_to_s_c(insert));
 444                 if (ret)
 445                         goto err;
 446
 447                 if (done)
 448                         goto out;
 449 next:
 450                 bch2_btree_iter_advance(&iter);
 451                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
 452                 if ((ret = bkey_err(k)))
 453                         goto err;
 454                 if (!k.k)
 455                         goto out;
 456         }
 457
 458         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
 459                 ret = extent_back_merge(trans, &iter, insert, k);
 460                 if (ret)
 461                         goto err;
 462         }
 463 out:
 464         if (!bkey_deleted(&insert->k))
 465                 ret = bch2_btree_insert_nonextent(trans, btree_id, insert, flags);
 466 err:
 467         bch2_trans_iter_exit(trans, &iter);
 468
 469         return ret;
 470 }
 471
 472 static noinline int flush_new_cached_update(struct btree_trans *trans,
 473                                             struct btree_insert_entry *i,
 474                                             enum btree_update_flags flags,
 475                                             unsigned long ip)
 476 {
 477         struct bkey k;
 478         int ret;
 479
 480         btree_path_idx_t path_idx =
 481                 bch2_path_get(trans, i->btree_id, i->old_k.p, 1, 0,
 482                               BTREE_ITER_INTENT, _THIS_IP_);
 483         ret = bch2_btree_path_traverse(trans, path_idx, 0);
 484         if (ret)
 485                 goto out;
 486
 487         struct btree_path *btree_path = trans->paths + path_idx;
 488
 489         /*
 490          * The old key in the insert entry might actually refer to an existing
 491          * key in the btree that has been deleted from cache and not yet
 492          * flushed. Check for this and skip the flush so we don't run triggers
 493          * against a stale key.
 494          */
 495         bch2_btree_path_peek_slot_exact(btree_path, &k);
 496         if (!bkey_deleted(&k))
 497                 goto out;
 498
 499         i->key_cache_already_flushed = true;
 500         i->flags |= BTREE_TRIGGER_NORUN;
 501
 502         btree_path_set_should_be_locked(btree_path);
 503         ret = bch2_trans_update_by_path(trans, path_idx, i->k, flags, ip);
 504 out:
 505         bch2_path_put(trans, path_idx, true);
 506         return ret;
 507 }
 508
 509 static int __must_check
 510 bch2_trans_update_by_path(struct btree_trans *trans, btree_path_idx_t path_idx,
 511                           struct bkey_i *k, enum btree_update_flags flags,
 512                           unsigned long ip)
 513 {
 514         struct bch_fs *c = trans->c;
 515         struct btree_insert_entry *i, n;
 516         int cmp;
 517
 518         struct btree_path *path = trans->paths + path_idx;
 519         EBUG_ON(!path->should_be_locked);
 520         EBUG_ON(trans->nr_updates >= trans->nr_paths);
 521         EBUG_ON(!bpos_eq(k->k.p, path->pos));
 522
 523         n = (struct btree_insert_entry) {
 524                 .flags          = flags,
 525                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
 526                 .btree_id       = path->btree_id,
 527                 .level          = path->level,
 528                 .cached         = path->cached,
 529                 .path           = path_idx,
 530                 .k              = k,
 531                 .ip_allocated   = ip,
 532         };
 533
 534 #ifdef CONFIG_BCACHEFS_DEBUG
 535         trans_for_each_update(trans, i)
 536                 BUG_ON(i != trans->updates &&
 537                        btree_insert_entry_cmp(i - 1, i) >= 0);
 538 #endif
 539
 540         /*
 541          * Pending updates are kept sorted: first, find position of new update,
 542          * then delete/trim any updates the new update overwrites:
 543          */
 544         for (i = trans->updates; i < trans->updates + trans->nr_updates; i++) {
 545                 cmp = btree_insert_entry_cmp(&n, i);
 546                 if (cmp <= 0)
 547                         break;
 548         }
 549
 550         if (!cmp && i < trans->updates + trans->nr_updates) {
 551                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
 552
 553                 bch2_path_put(trans, i->path, true);
 554                 i->flags        = n.flags;
 555                 i->cached       = n.cached;
 556                 i->k            = n.k;
 557                 i->path         = n.path;
 558                 i->ip_allocated = n.ip_allocated;
 559         } else {
 560                 array_insert_item(trans->updates, trans->nr_updates,
 561                                   i - trans->updates, n);
 562
 563                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
 564                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
 565
 566                 if (unlikely(trans->journal_replay_not_finished)) {
 567                         struct bkey_i *j_k =
 568                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
 569
 570                         if (j_k) {
 571                                 i->old_k = j_k->k;
 572                                 i->old_v = &j_k->v;
 573                         }
 574                 }
 575         }
 576
 577         __btree_path_get(trans->paths + i->path, true);
 578
 579         /*
 580          * If a key is present in the key cache, it must also exist in the
 581          * btree - this is necessary for cache coherency. When iterating over
 582          * a btree that's cached in the key cache, the btree iter code checks
 583          * the key cache - but the key has to exist in the btree for that to
 584          * work:
 585          */
 586         if (path->cached && bkey_deleted(&i->old_k))
 587                 return flush_new_cached_update(trans, i, flags, ip);
 588
 589         return 0;
 590 }
 591
 592 static noinline int bch2_trans_update_get_key_cache(struct btree_trans *trans,
 593                                                     struct btree_iter *iter,
 594                                                     struct btree_path *path)
 595 {
 596         struct btree_path *key_cache_path = btree_iter_key_cache_path(trans, iter);
 597
 598         if (!key_cache_path ||
 599             !key_cache_path->should_be_locked ||
 600             !bpos_eq(key_cache_path->pos, iter->pos)) {
 601                 struct bkey_cached *ck;
 602                 int ret;
 603
 604                 if (!iter->key_cache_path)
 605                         iter->key_cache_path =
 606                                 bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
 607                                               BTREE_ITER_INTENT|
 608                                               BTREE_ITER_CACHED, _THIS_IP_);
 609
 610                 iter->key_cache_path =
 611                         bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
 612                                                 iter->flags & BTREE_ITER_INTENT,
 613                                                 _THIS_IP_);
 614
 615                 ret = bch2_btree_path_traverse(trans, iter->key_cache_path, BTREE_ITER_CACHED);
 616                 if (unlikely(ret))
 617                         return ret;
 618
 619                 ck = (void *) trans->paths[iter->key_cache_path].l[0].b;
 620
 621                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 622                         trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
 623                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
 624                 }
 625
 626                 btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
 627         }
 628
 629         return 0;
 630 }
 631
 632 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 633                                    struct bkey_i *k, enum btree_update_flags flags)
 634 {
 635         if (iter->flags & BTREE_ITER_IS_EXTENTS)
 636                 return bch2_trans_update_extent(trans, iter, k, flags);
 637
 638         if (!(flags & (BTREE_UPDATE_SNAPSHOT_WHITEOUT_CHECKS_DONE|
 639                        BTREE_UPDATE_KEY_CACHE_RECLAIM)) &&
 640             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 641             bch2_snapshot_parent(trans->c, k->k.p.snapshot)) {
 642                 int ret = overwrite_interior_snapshot_key(trans, iter, k);
 643                 if (ret)
 644                         return ret < 0 ? ret : 0;
 645         }
 646
 647         /*
 648          * Ensure that updates to cached btrees go to the key cache:
 649          */
 650         btree_path_idx_t path_idx = iter->update_path ?: iter->path;
 651         struct btree_path *path = trans->paths + path_idx;
 652
 653         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
 654             !path->cached &&
 655             !path->level &&
 656             btree_id_cached(trans->c, path->btree_id)) {
 657                 int ret = bch2_trans_update_get_key_cache(trans, iter, path);
 658                 if (ret)
 659                         return ret;
 660
 661                 path_idx = iter->key_cache_path;
 662         }
 663
 664         return bch2_trans_update_by_path(trans, path_idx, k, flags, _RET_IP_);
 665 }
 666
 667 int bch2_btree_insert_clone_trans(struct btree_trans *trans,
 668                                   enum btree_id btree,
 669                                   struct bkey_i *k)
 670 {
 671         struct bkey_i *n = bch2_trans_kmalloc(trans, bkey_bytes(&k->k));
 672         int ret = PTR_ERR_OR_ZERO(n);
 673         if (ret)
 674                 return ret;
 675
 676         bkey_copy(n, k);
 677         return bch2_btree_insert_trans(trans, btree, n, 0);
 678 }
 679
 680 struct jset_entry *__bch2_trans_jset_entry_alloc(struct btree_trans *trans, unsigned u64s)
 681 {
 682         unsigned new_top = trans->journal_entries_u64s + u64s;
 683         unsigned old_size = trans->journal_entries_size;
 684
 685         if (new_top > trans->journal_entries_size) {
 686                 trans->journal_entries_size = roundup_pow_of_two(new_top);
 687
 688                 btree_trans_stats(trans)->journal_entries_size = trans->journal_entries_size;
 689         }
 690
 691         struct jset_entry *n =
 692                 bch2_trans_kmalloc_nomemzero(trans,
 693                                 trans->journal_entries_size * sizeof(u64));
 694         if (IS_ERR(n))
 695                 return ERR_CAST(n);
 696
 697         if (trans->journal_entries)
 698                 memcpy(n, trans->journal_entries, old_size * sizeof(u64));
 699         trans->journal_entries = n;
 700
 701         struct jset_entry *e = btree_trans_journal_entries_top(trans);
 702         trans->journal_entries_u64s = new_top;
 703         return e;
 704 }
 705
 706 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
 707                              enum btree_id btree, struct bpos end)
 708 {
 709         struct bkey_s_c k;
 710         int ret = 0;
 711
 712         bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
 713         k = bch2_btree_iter_prev(iter);
 714         ret = bkey_err(k);
 715         if (ret)
 716                 goto err;
 717
 718         bch2_btree_iter_advance(iter);
 719         k = bch2_btree_iter_peek_slot(iter);
 720         ret = bkey_err(k);
 721         if (ret)
 722                 goto err;
 723
 724         BUG_ON(k.k->type != KEY_TYPE_deleted);
 725
 726         if (bkey_gt(k.k->p, end)) {
 727                 ret = -BCH_ERR_ENOSPC_btree_slot;
 728                 goto err;
 729         }
 730
 731         return 0;
 732 err:
 733         bch2_trans_iter_exit(trans, iter);
 734         return ret;
 735 }
 736
 737 void bch2_trans_commit_hook(struct btree_trans *trans,
 738                             struct btree_trans_commit_hook *h)
 739 {
 740         h->next = trans->hooks;
 741         trans->hooks = h;
 742 }
 743
 744 int bch2_btree_insert_nonextent(struct btree_trans *trans,
 745                                 enum btree_id btree, struct bkey_i *k,
 746                                 enum btree_update_flags flags)
 747 {
 748         struct btree_iter iter;
 749         int ret;
 750
 751         bch2_trans_iter_init(trans, &iter, btree, k->k.p,
 752                              BTREE_ITER_CACHED|
 753                              BTREE_ITER_NOT_EXTENTS|
 754                              BTREE_ITER_INTENT);
 755         ret   = bch2_btree_iter_traverse(&iter) ?:
 756                 bch2_trans_update(trans, &iter, k, flags);
 757         bch2_trans_iter_exit(trans, &iter);
 758         return ret;
 759 }
 760
 761 int bch2_btree_insert_trans(struct btree_trans *trans, enum btree_id id,
 762                             struct bkey_i *k, enum btree_update_flags flags)
 763 {
 764         struct btree_iter iter;
 765         int ret;
 766
 767         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
 768                              BTREE_ITER_CACHED|
 769                              BTREE_ITER_INTENT);
 770         ret   = bch2_btree_iter_traverse(&iter) ?:
 771                 bch2_trans_update(trans, &iter, k, flags);
 772         bch2_trans_iter_exit(trans, &iter);
 773         return ret;
 774 }
 775
 776 /**
 777  * bch2_btree_insert - insert keys into the extent btree
 778  * @c:                  pointer to struct bch_fs
 779  * @id:                 btree to insert into
 780  * @k:                  key to insert
 781  * @disk_res:           must be non-NULL whenever inserting or potentially
 782  *                      splitting data extents
 783  * @flags:              transaction commit flags
 784  *
 785  * Returns:             0 on success, error code on failure
 786  */
 787 int bch2_btree_insert(struct bch_fs *c, enum btree_id id, struct bkey_i *k,
 788                       struct disk_reservation *disk_res, int flags)
 789 {
 790         return bch2_trans_do(c, disk_res, NULL, flags,
 791                              bch2_btree_insert_trans(trans, id, k, 0));
 792 }
 793
 794 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
 795                                 unsigned len, unsigned update_flags)
 796 {
 797         struct bkey_i *k;
 798
 799         k = bch2_trans_kmalloc(trans, sizeof(*k));
 800         if (IS_ERR(k))
 801                 return PTR_ERR(k);
 802
 803         bkey_init(&k->k);
 804         k->k.p = iter->pos;
 805         bch2_key_resize(&k->k, len);
 806         return bch2_trans_update(trans, iter, k, update_flags);
 807 }
 808
 809 int bch2_btree_delete_at(struct btree_trans *trans,
 810                          struct btree_iter *iter, unsigned update_flags)
 811 {
 812         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
 813 }
 814
 815 int bch2_btree_delete(struct btree_trans *trans,
 816                       enum btree_id btree, struct bpos pos,
 817                       unsigned update_flags)
 818 {
 819         struct btree_iter iter;
 820         int ret;
 821
 822         bch2_trans_iter_init(trans, &iter, btree, pos,
 823                              BTREE_ITER_CACHED|
 824                              BTREE_ITER_INTENT);
 825         ret   = bch2_btree_iter_traverse(&iter) ?:
 826                 bch2_btree_delete_at(trans, &iter, update_flags);
 827         bch2_trans_iter_exit(trans, &iter);
 828
 829         return ret;
 830 }
 831
 832 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
 833                                   struct bpos start, struct bpos end,
 834                                   unsigned update_flags,
 835                                   u64 *journal_seq)
 836 {
 837         u32 restart_count = trans->restart_count;
 838         struct btree_iter iter;
 839         struct bkey_s_c k;
 840         int ret = 0;
 841
 842         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
 843         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
 844                 struct disk_reservation disk_res =
 845                         bch2_disk_reservation_init(trans->c, 0);
 846                 struct bkey_i delete;
 847
 848                 ret = bkey_err(k);
 849                 if (ret)
 850                         goto err;
 851
 852                 bkey_init(&delete.k);
 853
 854                 /*
 855                  * This could probably be more efficient for extents:
 856                  */
 857
 858                 /*
 859                  * For extents, iter.pos won't necessarily be the same as
 860                  * bkey_start_pos(k.k) (for non extents they always will be the
 861                  * same). It's important that we delete starting from iter.pos
 862                  * because the range we want to delete could start in the middle
 863                  * of k.
 864                  *
 865                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 866                  * bkey_start_pos(k.k)).
 867                  */
 868                 delete.k.p = iter.pos;
 869
 870                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
 871                         bch2_key_resize(&delete.k,
 872                                         bpos_min(end, k.k->p).offset -
 873                                         iter.pos.offset);
 874
 875                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
 876                         bch2_trans_commit(trans, &disk_res, journal_seq,
 877                                           BCH_TRANS_COMMIT_no_enospc);
 878                 bch2_disk_reservation_put(trans->c, &disk_res);
 879 err:
 880                 /*
 881                  * the bch2_trans_begin() call is in a weird place because we
 882                  * need to call it after every transaction commit, to avoid path
 883                  * overflow, but don't want to call it if the delete operation
 884                  * is a no-op and we have no work to do:
 885                  */
 886                 bch2_trans_begin(trans);
 887
 888                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 889                         ret = 0;
 890                 if (ret)
 891                         break;
 892         }
 893         bch2_trans_iter_exit(trans, &iter);
 894
 895         return ret ?: trans_was_restarted(trans, restart_count);
 896 }
 897
 898 /*
 899  * bch_btree_delete_range - delete everything within a given range
 900  *
 901  * Range is a half open interval - [start, end)
 902  */
 903 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 904                             struct bpos start, struct bpos end,
 905                             unsigned update_flags,
 906                             u64 *journal_seq)
 907 {
 908         int ret = bch2_trans_run(c,
 909                         bch2_btree_delete_range_trans(trans, id, start, end,
 910                                                       update_flags, journal_seq));
 911         if (ret == -BCH_ERR_transaction_restart_nested)
 912                 ret = 0;
 913         return ret;
 914 }
 915
 916 int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
 917                        struct bpos pos, bool set)
 918 {
 919         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(*k));
 920         int ret = PTR_ERR_OR_ZERO(k);
 921         if (ret)
 922                 return ret;
 923
 924         bkey_init(&k->k);
 925         k->k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
 926         k->k.p = pos;
 927
 928         struct btree_iter iter;
 929         bch2_trans_iter_init(trans, &iter, btree, pos, BTREE_ITER_INTENT);
 930
 931         ret   = bch2_btree_iter_traverse(&iter) ?:
 932                 bch2_trans_update(trans, &iter, k, 0);
 933         bch2_trans_iter_exit(trans, &iter);
 934         return ret;
 935 }
 936
 937 int bch2_btree_bit_mod_buffered(struct btree_trans *trans, enum btree_id btree,
 938                                 struct bpos pos, bool set)
 939 {
 940         struct bkey_i k;
 941
 942         bkey_init(&k.k);
 943         k.k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
 944         k.k.p = pos;
 945
 946         return bch2_trans_update_buffered(trans, btree, &k);
 947 }
 948
 949 static int __bch2_trans_log_msg(struct btree_trans *trans, struct printbuf *buf, unsigned u64s)
 950 {
 951         struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, jset_u64s(u64s));
 952         int ret = PTR_ERR_OR_ZERO(e);
 953         if (ret)
 954                 return ret;
 955
 956         struct jset_entry_log *l = container_of(e, struct jset_entry_log, entry);
 957         journal_entry_init(e, BCH_JSET_ENTRY_log, 0, 1, u64s);
 958         memcpy(l->d, buf->buf, buf->pos);
 959         return 0;
 960 }
 961
 962 __printf(3, 0)
 963 static int
 964 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
 965                   va_list args)
 966 {
 967         struct printbuf buf = PRINTBUF;
 968         prt_vprintf(&buf, fmt, args);
 969
 970         unsigned u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
 971         prt_chars(&buf, '\0', u64s * sizeof(u64) - buf.pos);
 972
 973         int ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
 974         if (ret)
 975                 goto err;
 976
 977         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
 978                 ret = darray_make_room(&c->journal.early_journal_entries, jset_u64s(u64s));
 979                 if (ret)
 980                         goto err;
 981
 982                 struct jset_entry_log *l = (void *) &darray_top(c->journal.early_journal_entries);
 983                 journal_entry_init(&l->entry, BCH_JSET_ENTRY_log, 0, 1, u64s);
 984                 memcpy(l->d, buf.buf, buf.pos);
 985                 c->journal.early_journal_entries.nr += jset_u64s(u64s);
 986         } else {
 987                 ret = bch2_trans_do(c, NULL, NULL,
 988                         BCH_TRANS_COMMIT_lazy_rw|commit_flags,
 989                         __bch2_trans_log_msg(trans, &buf, u64s));
 990         }
 991 err:
 992         printbuf_exit(&buf);
 993         return ret;
 994 }
 995
 996 __printf(2, 3)
 997 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
 998 {
 999         va_list args;
1000         int ret;
1001
1002         va_start(args, fmt);
1003         ret = __bch2_fs_log_msg(c, 0, fmt, args);
1004         va_end(args);
1005         return ret;
1006 }
1007
1008 /*
1009  * Use for logging messages during recovery to enable reserved space and avoid
1010  * blocking.
1011  */
1012 __printf(2, 3)
1013 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
1014 {
1015         va_list args;
1016         int ret;
1017
1018         va_start(args, fmt);
1019         ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);
1020         va_end(args);
1021         return ret;
1022 }