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

   1
   2 #include "bcachefs.h"
   3 #include "btree_update.h"
   4 #include "btree_update_interior.h"
   5 #include "btree_io.h"
   6 #include "btree_iter.h"
   7 #include "btree_locking.h"
   8 #include "debug.h"
   9 #include "extents.h"
  10 #include "journal.h"
  11 #include "keylist.h"
  12
  13 #include <linux/sort.h>
  14 #include <trace/events/bcachefs.h>
  15
  16 /* Inserting into a given leaf node (last stage of insert): */
  17
  18 /* Handle overwrites and do insert, for non extents: */
  19 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  20                                 struct btree *b,
  21                                 struct btree_node_iter *node_iter,
  22                                 struct bkey_i *insert)
  23 {
  24         const struct bkey_format *f = &b->format;
  25         struct bkey_packed *k;
  26         struct bset_tree *t;
  27         unsigned clobber_u64s;
  28
  29         EBUG_ON(btree_node_just_written(b));
  30         EBUG_ON(bset_written(b, btree_bset_last(b)));
  31         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  32         EBUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0 ||
  33                 bkey_cmp(insert->k.p, b->data->max_key) > 0);
  34
  35         k = bch2_btree_node_iter_peek_all(node_iter, b);
  36         if (k && !bkey_cmp_packed(b, k, &insert->k)) {
  37                 BUG_ON(bkey_whiteout(k));
  38
  39                 t = bch2_bkey_to_bset(b, k);
  40
  41                 if (bset_unwritten(b, bset(b, t)) &&
  42                     bkey_val_u64s(&insert->k) == bkeyp_val_u64s(f, k) &&
  43                     !bkey_whiteout(&insert->k)) {
  44                         k->type = insert->k.type;
  45                         memcpy_u64s(bkeyp_val(f, k), &insert->v,
  46                                     bkey_val_u64s(&insert->k));
  47                         return true;
  48                 }
  49
  50                 insert->k.needs_whiteout = k->needs_whiteout;
  51
  52                 btree_keys_account_key_drop(&b->nr, t - b->set, k);
  53
  54                 if (t == bset_tree_last(b)) {
  55                         clobber_u64s = k->u64s;
  56
  57                         /*
  58                          * If we're deleting, and the key we're deleting doesn't
  59                          * need a whiteout (it wasn't overwriting a key that had
  60                          * been written to disk) - just delete it:
  61                          */
  62                         if (bkey_whiteout(&insert->k) && !k->needs_whiteout) {
  63                                 bch2_bset_delete(b, k, clobber_u64s);
  64                                 bch2_btree_node_iter_fix(iter, b, node_iter, t,
  65                                                         k, clobber_u64s, 0);
  66                                 return true;
  67                         }
  68
  69                         goto overwrite;
  70                 }
  71
  72                 k->type = KEY_TYPE_DELETED;
  73                 bch2_btree_node_iter_fix(iter, b, node_iter, t, k,
  74                                         k->u64s, k->u64s);
  75
  76                 if (bkey_whiteout(&insert->k)) {
  77                         reserve_whiteout(b, t, k);
  78                         return true;
  79                 } else {
  80                         k->needs_whiteout = false;
  81                 }
  82         } else {
  83                 /*
  84                  * Deleting, but the key to delete wasn't found - nothing to do:
  85                  */
  86                 if (bkey_whiteout(&insert->k))
  87                         return false;
  88
  89                 insert->k.needs_whiteout = false;
  90         }
  91
  92         t = bset_tree_last(b);
  93         k = bch2_btree_node_iter_bset_pos(node_iter, b, t);
  94         clobber_u64s = 0;
  95 overwrite:
  96         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
  97         if (k->u64s != clobber_u64s || bkey_whiteout(&insert->k))
  98                 bch2_btree_node_iter_fix(iter, b, node_iter, t, k,
  99                                         clobber_u64s, k->u64s);
 100         return true;
 101 }
 102
 103 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 104                                unsigned i, u64 seq)
 105 {
 106         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 107         struct btree_write *w = container_of(pin, struct btree_write, journal);
 108         struct btree *b = container_of(w, struct btree, writes[i]);
 109
 110         six_lock_read(&b->lock);
 111         bch2_btree_node_write_cond(c, b,
 112                         (btree_current_write(b) == w &&
 113                          w->journal.pin_list == journal_seq_pin(j, seq)));
 114         six_unlock_read(&b->lock);
 115 }
 116
 117 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 118 {
 119         return __btree_node_flush(j, pin, 0, seq);
 120 }
 121
 122 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 123 {
 124         return __btree_node_flush(j, pin, 1, seq);
 125 }
 126
 127 void bch2_btree_journal_key(struct btree_insert *trans,
 128                            struct btree_iter *iter,
 129                            struct bkey_i *insert)
 130 {
 131         struct bch_fs *c = trans->c;
 132         struct journal *j = &c->journal;
 133         struct btree *b = iter->l[0].b;
 134         struct btree_write *w = btree_current_write(b);
 135
 136         EBUG_ON(iter->level || b->level);
 137         EBUG_ON(trans->journal_res.ref !=
 138                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 139
 140         if (likely(trans->journal_res.ref)) {
 141                 u64 seq = trans->journal_res.seq;
 142                 bool needs_whiteout = insert->k.needs_whiteout;
 143
 144                 /* ick */
 145                 insert->k.needs_whiteout = false;
 146                 bch2_journal_add_keys(j, &trans->journal_res,
 147                                       iter->btree_id, insert);
 148                 insert->k.needs_whiteout = needs_whiteout;
 149
 150                 bch2_journal_set_has_inode(j, &trans->journal_res,
 151                                            insert->k.p.inode);
 152
 153                 if (trans->journal_seq)
 154                         *trans->journal_seq = seq;
 155                 btree_bset_last(b)->journal_seq = cpu_to_le64(seq);
 156         }
 157
 158         if (unlikely(!journal_pin_active(&w->journal)))
 159                 bch2_journal_pin_add(j, &trans->journal_res,
 160                                      &w->journal,
 161                                      btree_node_write_idx(b) == 0
 162                                      ? btree_node_flush0
 163                                      : btree_node_flush1);
 164
 165         if (unlikely(!btree_node_dirty(b)))
 166                 set_btree_node_dirty(b);
 167 }
 168
 169 static enum btree_insert_ret
 170 bch2_insert_fixup_key(struct btree_insert *trans,
 171                      struct btree_insert_entry *insert)
 172 {
 173         struct btree_iter *iter = insert->iter;
 174         struct btree_iter_level *l = &iter->l[0];
 175
 176         EBUG_ON(iter->level);
 177         EBUG_ON(insert->k->k.u64s >
 178                 bch_btree_keys_u64s_remaining(trans->c, l->b));
 179
 180         if (bch2_btree_bset_insert_key(iter, l->b, &l->iter,
 181                                        insert->k))
 182                 bch2_btree_journal_key(trans, iter, insert->k);
 183
 184         trans->did_work = true;
 185         return BTREE_INSERT_OK;
 186 }
 187
 188 /**
 189  * btree_insert_key - insert a key one key into a leaf node
 190  */
 191 static enum btree_insert_ret
 192 btree_insert_key_leaf(struct btree_insert *trans,
 193                       struct btree_insert_entry *insert)
 194 {
 195         struct bch_fs *c = trans->c;
 196         struct btree_iter *iter = insert->iter;
 197         struct btree *b = iter->l[0].b;
 198         enum btree_insert_ret ret;
 199         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
 200         int old_live_u64s = b->nr.live_u64s;
 201         int live_u64s_added, u64s_added;
 202
 203         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 204
 205         ret = !btree_node_is_extents(b)
 206                 ? bch2_insert_fixup_key(trans, insert)
 207                 : bch2_insert_fixup_extent(trans, insert);
 208
 209         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 210         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
 211
 212         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 213                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 214         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 215                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 216
 217         if (u64s_added > live_u64s_added &&
 218             bch2_maybe_compact_whiteouts(c, b))
 219                 bch2_btree_iter_reinit_node(iter, b);
 220
 221         trace_btree_insert_key(c, b, insert->k);
 222         return ret;
 223 }
 224
 225 static bool same_leaf_as_prev(struct btree_insert *trans,
 226                               struct btree_insert_entry *i)
 227 {
 228         /*
 229          * Because we sorted the transaction entries, if multiple iterators
 230          * point to the same leaf node they'll always be adjacent now:
 231          */
 232         return i != trans->entries &&
 233                 i[0].iter->l[0].b == i[-1].iter->l[0].b;
 234 }
 235
 236 #define trans_for_each_entry(trans, i)                                  \
 237         for ((i) = (trans)->entries; (i) < (trans)->entries + (trans)->nr; (i)++)
 238
 239 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
 240                                             struct btree_iter *iter)
 241 {
 242         bch2_btree_node_lock_write(b, iter);
 243
 244         if (btree_node_just_written(b) &&
 245             bch2_btree_post_write_cleanup(c, b))
 246                 bch2_btree_iter_reinit_node(iter, b);
 247
 248         /*
 249          * If the last bset has been written, or if it's gotten too big - start
 250          * a new bset to insert into:
 251          */
 252         if (want_new_bset(c, b))
 253                 bch2_btree_init_next(c, b, iter);
 254 }
 255
 256 static void multi_lock_write(struct bch_fs *c, struct btree_insert *trans)
 257 {
 258         struct btree_insert_entry *i;
 259
 260         trans_for_each_entry(trans, i)
 261                 if (!same_leaf_as_prev(trans, i))
 262                         bch2_btree_node_lock_for_insert(c, i->iter->l[0].b,
 263                                                         i->iter);
 264 }
 265
 266 static void multi_unlock_write(struct btree_insert *trans)
 267 {
 268         struct btree_insert_entry *i;
 269
 270         trans_for_each_entry(trans, i)
 271                 if (!same_leaf_as_prev(trans, i))
 272                         bch2_btree_node_unlock_write(i->iter->l[0].b, i->iter);
 273 }
 274
 275 static inline void btree_trans_sort(struct btree_insert *trans)
 276 {
 277         int i, end = trans->nr;
 278
 279         while (--end > 0)
 280                 for (i = 0; i < end; i++)
 281                         if (btree_iter_cmp(trans->entries[i].iter,
 282                                            trans->entries[i + 1].iter) > 0)
 283                                 swap(trans->entries[i], trans->entries[i + 1]);
 284 }
 285
 286 /* Normal update interface: */
 287
 288 /**
 289  * __bch_btree_insert_at - insert keys at given iterator positions
 290  *
 291  * This is main entry point for btree updates.
 292  *
 293  * Return values:
 294  * -EINTR: locking changed, this function should be called again. Only returned
 295  *  if passed BTREE_INSERT_ATOMIC.
 296  * -EROFS: filesystem read only
 297  * -EIO: journal or btree node IO error
 298  */
 299 int __bch2_btree_insert_at(struct btree_insert *trans)
 300 {
 301         struct bch_fs *c = trans->c;
 302         struct btree_insert_entry *i;
 303         struct btree_iter *split = NULL;
 304         bool cycle_gc_lock = false;
 305         unsigned u64s;
 306         int ret;
 307
 308         trans_for_each_entry(trans, i) {
 309                 BUG_ON(i->iter->level);
 310                 BUG_ON(bkey_cmp(bkey_start_pos(&i->k->k), i->iter->pos));
 311                 BUG_ON(debug_check_bkeys(c) &&
 312                        bch2_bkey_invalid(c, i->iter->btree_id,
 313                                          bkey_i_to_s_c(i->k)));
 314         }
 315
 316         btree_trans_sort(trans);
 317
 318         if (unlikely(!percpu_ref_tryget(&c->writes)))
 319                 return -EROFS;
 320 retry_locks:
 321         ret = -EINTR;
 322         trans_for_each_entry(trans, i) {
 323                 if (!bch2_btree_iter_set_locks_want(i->iter, 1))
 324                         goto err;
 325
 326                 if (i->iter->uptodate == BTREE_ITER_NEED_TRAVERSE) {
 327                         ret = bch2_btree_iter_traverse(i->iter);
 328                         if (ret)
 329                                 goto err;
 330                 }
 331         }
 332 retry:
 333         trans->did_work = false;
 334         u64s = 0;
 335         trans_for_each_entry(trans, i)
 336                 if (!i->done)
 337                         u64s += jset_u64s(i->k->k.u64s + i->extra_res);
 338
 339         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 340
 341         ret = !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)
 342                 ? bch2_journal_res_get(&c->journal,
 343                                       &trans->journal_res,
 344                                       u64s, u64s)
 345                 : 0;
 346         if (ret)
 347                 goto err;
 348
 349         multi_lock_write(c, trans);
 350
 351         if (race_fault()) {
 352                 ret = -EINTR;
 353                 goto unlock;
 354         }
 355
 356         u64s = 0;
 357         trans_for_each_entry(trans, i) {
 358                 /* Multiple inserts might go to same leaf: */
 359                 if (!same_leaf_as_prev(trans, i))
 360                         u64s = 0;
 361
 362                 /*
 363                  * bch2_btree_node_insert_fits() must be called under write lock:
 364                  * with only an intent lock, another thread can still call
 365                  * bch2_btree_node_write(), converting an unwritten bset to a
 366                  * written one
 367                  */
 368                 if (!i->done) {
 369                         u64s += i->k->k.u64s + i->extra_res;
 370                         if (!bch2_btree_node_insert_fits(c,
 371                                         i->iter->l[0].b, u64s)) {
 372                                 split = i->iter;
 373                                 goto unlock;
 374                         }
 375                 }
 376         }
 377
 378         ret = 0;
 379         split = NULL;
 380         cycle_gc_lock = false;
 381
 382         trans_for_each_entry(trans, i) {
 383                 if (i->done)
 384                         continue;
 385
 386                 switch (btree_insert_key_leaf(trans, i)) {
 387                 case BTREE_INSERT_OK:
 388                         i->done = true;
 389                         break;
 390                 case BTREE_INSERT_JOURNAL_RES_FULL:
 391                 case BTREE_INSERT_NEED_TRAVERSE:
 392                         ret = -EINTR;
 393                         break;
 394                 case BTREE_INSERT_NEED_RESCHED:
 395                         ret = -EAGAIN;
 396                         break;
 397                 case BTREE_INSERT_BTREE_NODE_FULL:
 398                         split = i->iter;
 399                         break;
 400                 case BTREE_INSERT_ENOSPC:
 401                         ret = -ENOSPC;
 402                         break;
 403                 case BTREE_INSERT_NEED_GC_LOCK:
 404                         cycle_gc_lock = true;
 405                         ret = -EINTR;
 406                         break;
 407                 default:
 408                         BUG();
 409                 }
 410
 411                 if (!trans->did_work && (ret || split))
 412                         break;
 413         }
 414 unlock:
 415         multi_unlock_write(trans);
 416         bch2_journal_res_put(&c->journal, &trans->journal_res);
 417
 418         if (split)
 419                 goto split;
 420         if (ret)
 421                 goto err;
 422
 423         trans_for_each_entry(trans, i)
 424                 if (i->iter->flags & BTREE_ITER_AT_END_OF_LEAF)
 425                         goto out;
 426
 427         trans_for_each_entry(trans, i) {
 428                 /*
 429                  * iterators are inconsistent when they hit end of leaf, until
 430                  * traversed again
 431                  */
 432                 if (i->iter->uptodate < BTREE_ITER_NEED_TRAVERSE &&
 433                     !same_leaf_as_prev(trans, i))
 434                         bch2_foreground_maybe_merge(c, i->iter, 0);
 435         }
 436 out:
 437         /* make sure we didn't lose an error: */
 438         if (!ret && IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 439                 trans_for_each_entry(trans, i)
 440                         BUG_ON(!i->done);
 441
 442         percpu_ref_put(&c->writes);
 443         return ret;
 444 split:
 445         /*
 446          * have to drop journal res before splitting, because splitting means
 447          * allocating new btree nodes, and holding a journal reservation
 448          * potentially blocks the allocator:
 449          */
 450         ret = bch2_btree_split_leaf(c, split, trans->flags);
 451         if (ret)
 452                 goto err;
 453         /*
 454          * if the split didn't have to drop locks the insert will still be
 455          * atomic (in the BTREE_INSERT_ATOMIC sense, what the caller peeked()
 456          * and is overwriting won't have changed)
 457          */
 458         goto retry_locks;
 459 err:
 460         if (cycle_gc_lock) {
 461                 down_read(&c->gc_lock);
 462                 up_read(&c->gc_lock);
 463         }
 464
 465         if (ret == -EINTR) {
 466                 trans_for_each_entry(trans, i) {
 467                         int ret2 = bch2_btree_iter_traverse(i->iter);
 468                         if (ret2) {
 469                                 ret = ret2;
 470                                 goto out;
 471                         }
 472                 }
 473
 474                 /*
 475                  * BTREE_ITER_ATOMIC means we have to return -EINTR if we
 476                  * dropped locks:
 477                  */
 478                 if (!(trans->flags & BTREE_INSERT_ATOMIC))
 479                         goto retry;
 480         }
 481
 482         goto out;
 483 }
 484
 485 int bch2_btree_delete_at(struct btree_iter *iter, unsigned flags)
 486 {
 487         struct bkey_i k;
 488
 489         bkey_init(&k.k);
 490         k.k.p = iter->pos;
 491
 492         return bch2_btree_insert_at(iter->c, NULL, NULL, NULL,
 493                                     BTREE_INSERT_NOFAIL|
 494                                     BTREE_INSERT_USE_RESERVE|flags,
 495                                     BTREE_INSERT_ENTRY(iter, &k));
 496 }
 497
 498 int bch2_btree_insert_list_at(struct btree_iter *iter,
 499                              struct keylist *keys,
 500                              struct disk_reservation *disk_res,
 501                              struct extent_insert_hook *hook,
 502                              u64 *journal_seq, unsigned flags)
 503 {
 504         BUG_ON(flags & BTREE_INSERT_ATOMIC);
 505         BUG_ON(bch2_keylist_empty(keys));
 506         bch2_verify_keylist_sorted(keys);
 507
 508         while (!bch2_keylist_empty(keys)) {
 509                 int ret = bch2_btree_insert_at(iter->c, disk_res, hook,
 510                                 journal_seq, flags,
 511                                 BTREE_INSERT_ENTRY(iter, bch2_keylist_front(keys)));
 512                 if (ret)
 513                         return ret;
 514
 515                 bch2_keylist_pop_front(keys);
 516         }
 517
 518         return 0;
 519 }
 520
 521 /**
 522  * bch_btree_insert - insert keys into the extent btree
 523  * @c:                  pointer to struct bch_fs
 524  * @id:                 btree to insert into
 525  * @insert_keys:        list of keys to insert
 526  * @hook:               insert callback
 527  */
 528 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
 529                      struct bkey_i *k,
 530                      struct disk_reservation *disk_res,
 531                      struct extent_insert_hook *hook,
 532                      u64 *journal_seq, int flags)
 533 {
 534         struct btree_iter iter;
 535         int ret;
 536
 537         bch2_btree_iter_init(&iter, c, id, bkey_start_pos(&k->k),
 538                              BTREE_ITER_INTENT);
 539         ret = bch2_btree_insert_at(c, disk_res, hook, journal_seq, flags,
 540                                   BTREE_INSERT_ENTRY(&iter, k));
 541         bch2_btree_iter_unlock(&iter);
 542
 543         return ret;
 544 }
 545
 546 /*
 547  * bch_btree_delete_range - delete everything within a given range
 548  *
 549  * Range is a half open interval - [start, end)
 550  */
 551 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 552                            struct bpos start,
 553                            struct bpos end,
 554                            struct bversion version,
 555                            struct disk_reservation *disk_res,
 556                            struct extent_insert_hook *hook,
 557                            u64 *journal_seq)
 558 {
 559         struct btree_iter iter;
 560         struct bkey_s_c k;
 561         int ret = 0;
 562
 563         bch2_btree_iter_init(&iter, c, id, start,
 564                              BTREE_ITER_INTENT);
 565
 566         while ((k = bch2_btree_iter_peek(&iter)).k &&
 567                !(ret = btree_iter_err(k))) {
 568                 unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits);
 569                 /* really shouldn't be using a bare, unpadded bkey_i */
 570                 struct bkey_i delete;
 571
 572                 if (bkey_cmp(iter.pos, end) >= 0)
 573                         break;
 574
 575                 bkey_init(&delete.k);
 576
 577                 /*
 578                  * For extents, iter.pos won't necessarily be the same as
 579                  * bkey_start_pos(k.k) (for non extents they always will be the
 580                  * same). It's important that we delete starting from iter.pos
 581                  * because the range we want to delete could start in the middle
 582                  * of k.
 583                  *
 584                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 585                  * bkey_start_pos(k.k)).
 586                  */
 587                 delete.k.p = iter.pos;
 588                 delete.k.version = version;
 589
 590                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
 591                         /*
 592                          * The extents btree is special - KEY_TYPE_DISCARD is
 593                          * used for deletions, not KEY_TYPE_DELETED. This is an
 594                          * internal implementation detail that probably
 595                          * shouldn't be exposed (internally, KEY_TYPE_DELETED is
 596                          * used as a proxy for k->size == 0):
 597                          */
 598                         delete.k.type = KEY_TYPE_DISCARD;
 599
 600                         /* create the biggest key we can */
 601                         bch2_key_resize(&delete.k, max_sectors);
 602                         bch2_cut_back(end, &delete.k);
 603                 }
 604
 605                 ret = bch2_btree_insert_at(c, disk_res, hook, journal_seq,
 606                                           BTREE_INSERT_NOFAIL,
 607                                           BTREE_INSERT_ENTRY(&iter, &delete));
 608                 if (ret)
 609                         break;
 610
 611                 bch2_btree_iter_cond_resched(&iter);
 612         }
 613
 614         bch2_btree_iter_unlock(&iter);
 615         return ret;
 616 }