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