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 "journal_reclaim.h"
  12 #include "keylist.h"
  13
  14 #include <linux/sort.h>
  15 #include <trace/events/bcachefs.h>
  16
  17 /* Inserting into a given leaf node (last stage of insert): */
  18
  19 /* Handle overwrites and do insert, for non extents: */
  20 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  21                                 struct btree *b,
  22                                 struct btree_node_iter *node_iter,
  23                                 struct bkey_i *insert)
  24 {
  25         const struct bkey_format *f = &b->format;
  26         struct bkey_packed *k;
  27         struct bset_tree *t;
  28         unsigned clobber_u64s;
  29
  30         EBUG_ON(btree_node_just_written(b));
  31         EBUG_ON(bset_written(b, btree_bset_last(b)));
  32         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  33         EBUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0 ||
  34                 bkey_cmp(insert->k.p, b->data->max_key) > 0);
  35
  36         k = bch2_btree_node_iter_peek_all(node_iter, b);
  37         if (k && !bkey_cmp_packed(b, k, &insert->k)) {
  38                 BUG_ON(bkey_whiteout(k));
  39
  40                 t = bch2_bkey_to_bset(b, k);
  41
  42                 if (bset_unwritten(b, bset(b, t)) &&
  43                     bkey_val_u64s(&insert->k) == bkeyp_val_u64s(f, k) &&
  44                     !bkey_whiteout(&insert->k)) {
  45                         k->type = insert->k.type;
  46                         memcpy_u64s(bkeyp_val(f, k), &insert->v,
  47                                     bkey_val_u64s(&insert->k));
  48                         return true;
  49                 }
  50
  51                 insert->k.needs_whiteout = k->needs_whiteout;
  52
  53                 btree_keys_account_key_drop(&b->nr, t - b->set, k);
  54
  55                 if (t == bset_tree_last(b)) {
  56                         clobber_u64s = k->u64s;
  57
  58                         /*
  59                          * If we're deleting, and the key we're deleting doesn't
  60                          * need a whiteout (it wasn't overwriting a key that had
  61                          * been written to disk) - just delete it:
  62                          */
  63                         if (bkey_whiteout(&insert->k) && !k->needs_whiteout) {
  64                                 bch2_bset_delete(b, k, clobber_u64s);
  65                                 bch2_btree_node_iter_fix(iter, b, node_iter, t,
  66                                                         k, clobber_u64s, 0);
  67                                 return true;
  68                         }
  69
  70                         goto overwrite;
  71                 }
  72
  73                 k->type = KEY_TYPE_DELETED;
  74                 bch2_btree_node_iter_fix(iter, b, node_iter, t, k,
  75                                         k->u64s, k->u64s);
  76
  77                 if (bkey_whiteout(&insert->k)) {
  78                         reserve_whiteout(b, t, k);
  79                         return true;
  80                 } else {
  81                         k->needs_whiteout = false;
  82                 }
  83         } else {
  84                 /*
  85                  * Deleting, but the key to delete wasn't found - nothing to do:
  86                  */
  87                 if (bkey_whiteout(&insert->k))
  88                         return false;
  89
  90                 insert->k.needs_whiteout = false;
  91         }
  92
  93         t = bset_tree_last(b);
  94         k = bch2_btree_node_iter_bset_pos(node_iter, b, t);
  95         clobber_u64s = 0;
  96 overwrite:
  97         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
  98         if (k->u64s != clobber_u64s || bkey_whiteout(&insert->k))
  99                 bch2_btree_node_iter_fix(iter, b, node_iter, t, k,
 100                                         clobber_u64s, k->u64s);
 101         return true;
 102 }
 103
 104 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 105                                unsigned i, u64 seq)
 106 {
 107         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 108         struct btree_write *w = container_of(pin, struct btree_write, journal);
 109         struct btree *b = container_of(w, struct btree, writes[i]);
 110
 111         btree_node_lock_type(c, b, SIX_LOCK_read);
 112         bch2_btree_node_write_cond(c, b,
 113                         (btree_current_write(b) == w &&
 114                          w->journal.pin_list == journal_seq_pin(j, seq)));
 115         six_unlock_read(&b->lock);
 116 }
 117
 118 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 119 {
 120         return __btree_node_flush(j, pin, 0, seq);
 121 }
 122
 123 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 124 {
 125         return __btree_node_flush(j, pin, 1, seq);
 126 }
 127
 128 void bch2_btree_journal_key(struct btree_insert *trans,
 129                            struct btree_iter *iter,
 130                            struct bkey_i *insert)
 131 {
 132         struct bch_fs *c = trans->c;
 133         struct journal *j = &c->journal;
 134         struct btree *b = iter->l[0].b;
 135         struct btree_write *w = btree_current_write(b);
 136
 137         EBUG_ON(iter->level || b->level);
 138         EBUG_ON(trans->journal_res.ref !=
 139                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 140
 141         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 142                 u64 seq = trans->journal_res.seq;
 143                 bool needs_whiteout = insert->k.needs_whiteout;
 144
 145                 /* ick */
 146                 insert->k.needs_whiteout = false;
 147                 bch2_journal_add_keys(j, &trans->journal_res,
 148                                       iter->btree_id, insert);
 149                 insert->k.needs_whiteout = needs_whiteout;
 150
 151                 bch2_journal_set_has_inode(j, &trans->journal_res,
 152                                            insert->k.p.inode);
 153
 154                 if (trans->journal_seq)
 155                         *trans->journal_seq = seq;
 156                 btree_bset_last(b)->journal_seq = cpu_to_le64(seq);
 157         }
 158
 159         if (unlikely(!journal_pin_active(&w->journal))) {
 160                 u64 seq = likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
 161                         ? trans->journal_res.seq
 162                         : j->replay_journal_seq;
 163
 164                 bch2_journal_pin_add(j, seq, &w->journal,
 165                                      btree_node_write_idx(b) == 0
 166                                      ? btree_node_flush0
 167                                      : btree_node_flush1);
 168         }
 169
 170         if (unlikely(!btree_node_dirty(b)))
 171                 set_btree_node_dirty(b);
 172 }
 173
 174 static enum btree_insert_ret
 175 bch2_insert_fixup_key(struct btree_insert *trans,
 176                      struct btree_insert_entry *insert)
 177 {
 178         struct btree_iter *iter = insert->iter;
 179         struct btree_iter_level *l = &iter->l[0];
 180
 181         EBUG_ON(iter->level);
 182         EBUG_ON(insert->k->k.u64s >
 183                 bch_btree_keys_u64s_remaining(trans->c, l->b));
 184
 185         if (bch2_btree_bset_insert_key(iter, l->b, &l->iter,
 186                                        insert->k))
 187                 bch2_btree_journal_key(trans, iter, insert->k);
 188
 189         trans->did_work = true;
 190         return BTREE_INSERT_OK;
 191 }
 192
 193 /**
 194  * btree_insert_key - insert a key one key into a leaf node
 195  */
 196 static enum btree_insert_ret
 197 btree_insert_key_leaf(struct btree_insert *trans,
 198                       struct btree_insert_entry *insert)
 199 {
 200         struct bch_fs *c = trans->c;
 201         struct btree_iter *iter = insert->iter;
 202         struct btree *b = iter->l[0].b;
 203         enum btree_insert_ret ret;
 204         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
 205         int old_live_u64s = b->nr.live_u64s;
 206         int live_u64s_added, u64s_added;
 207
 208         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 209
 210         ret = !btree_node_is_extents(b)
 211                 ? bch2_insert_fixup_key(trans, insert)
 212                 : bch2_insert_fixup_extent(trans, insert);
 213
 214         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 215         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
 216
 217         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 218                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 219         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 220                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 221
 222         if (u64s_added > live_u64s_added &&
 223             bch2_maybe_compact_whiteouts(c, b))
 224                 bch2_btree_iter_reinit_node(iter, b);
 225
 226         trace_btree_insert_key(c, b, insert->k);
 227         return ret;
 228 }
 229
 230 static bool same_leaf_as_prev(struct btree_insert *trans,
 231                               struct btree_insert_entry *i)
 232 {
 233         /*
 234          * Because we sorted the transaction entries, if multiple iterators
 235          * point to the same leaf node they'll always be adjacent now:
 236          */
 237         return i != trans->entries &&
 238                 i[0].iter->l[0].b == i[-1].iter->l[0].b;
 239 }
 240
 241 #define trans_for_each_entry(trans, i)                                  \
 242         for ((i) = (trans)->entries; (i) < (trans)->entries + (trans)->nr; (i)++)
 243
 244 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
 245                                             struct btree_iter *iter)
 246 {
 247         bch2_btree_node_lock_write(b, iter);
 248
 249         if (btree_node_just_written(b) &&
 250             bch2_btree_post_write_cleanup(c, b))
 251                 bch2_btree_iter_reinit_node(iter, b);
 252
 253         /*
 254          * If the last bset has been written, or if it's gotten too big - start
 255          * a new bset to insert into:
 256          */
 257         if (want_new_bset(c, b))
 258                 bch2_btree_init_next(c, b, iter);
 259 }
 260
 261 static void multi_lock_write(struct bch_fs *c, struct btree_insert *trans)
 262 {
 263         struct btree_insert_entry *i;
 264
 265         trans_for_each_entry(trans, i)
 266                 if (!same_leaf_as_prev(trans, i))
 267                         bch2_btree_node_lock_for_insert(c, i->iter->l[0].b,
 268                                                         i->iter);
 269 }
 270
 271 static void multi_unlock_write(struct btree_insert *trans)
 272 {
 273         struct btree_insert_entry *i;
 274
 275         trans_for_each_entry(trans, i)
 276                 if (!same_leaf_as_prev(trans, i))
 277                         bch2_btree_node_unlock_write(i->iter->l[0].b, i->iter);
 278 }
 279
 280 static inline int btree_trans_cmp(struct btree_insert_entry l,
 281                                   struct btree_insert_entry r)
 282 {
 283         return btree_iter_cmp(l.iter, r.iter);
 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         bubble_sort(trans->entries, trans->nr, btree_trans_cmp);
 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
 452         /*
 453          * This can happen when we insert part of an extent - with an update
 454          * with multiple keys, we don't want to redo the entire update - that's
 455          * just too confusing:
 456          */
 457         if (!ret &&
 458             (trans->flags & BTREE_INSERT_ATOMIC) &&
 459             trans->did_work)
 460                 ret = -EINTR;
 461
 462         if (ret)
 463                 goto err;
 464
 465         /*
 466          * if the split didn't have to drop locks the insert will still be
 467          * atomic (in the BTREE_INSERT_ATOMIC sense, what the caller peeked()
 468          * and is overwriting won't have changed)
 469          */
 470         goto retry_locks;
 471 err:
 472         if (cycle_gc_lock) {
 473                 down_read(&c->gc_lock);
 474                 up_read(&c->gc_lock);
 475         }
 476
 477         if (ret == -EINTR) {
 478                 trans_for_each_entry(trans, i) {
 479                         int ret2 = bch2_btree_iter_traverse(i->iter);
 480                         if (ret2) {
 481                                 ret = ret2;
 482                                 goto out;
 483                         }
 484                 }
 485
 486                 /*
 487                  * BTREE_ITER_ATOMIC means we have to return -EINTR if we
 488                  * dropped locks:
 489                  */
 490                 if (!(trans->flags & BTREE_INSERT_ATOMIC))
 491                         goto retry;
 492         }
 493
 494         goto out;
 495 }
 496
 497 int bch2_btree_delete_at(struct btree_iter *iter, unsigned flags)
 498 {
 499         struct bkey_i k;
 500
 501         bkey_init(&k.k);
 502         k.k.p = iter->pos;
 503
 504         return bch2_btree_insert_at(iter->c, NULL, NULL, NULL,
 505                                     BTREE_INSERT_NOFAIL|
 506                                     BTREE_INSERT_USE_RESERVE|flags,
 507                                     BTREE_INSERT_ENTRY(iter, &k));
 508 }
 509
 510 int bch2_btree_insert_list_at(struct btree_iter *iter,
 511                              struct keylist *keys,
 512                              struct disk_reservation *disk_res,
 513                              struct extent_insert_hook *hook,
 514                              u64 *journal_seq, unsigned flags)
 515 {
 516         BUG_ON(flags & BTREE_INSERT_ATOMIC);
 517         BUG_ON(bch2_keylist_empty(keys));
 518         bch2_verify_keylist_sorted(keys);
 519
 520         while (!bch2_keylist_empty(keys)) {
 521                 int ret = bch2_btree_insert_at(iter->c, disk_res, hook,
 522                                 journal_seq, flags,
 523                                 BTREE_INSERT_ENTRY(iter, bch2_keylist_front(keys)));
 524                 if (ret)
 525                         return ret;
 526
 527                 bch2_keylist_pop_front(keys);
 528         }
 529
 530         return 0;
 531 }
 532
 533 /**
 534  * bch_btree_insert - insert keys into the extent btree
 535  * @c:                  pointer to struct bch_fs
 536  * @id:                 btree to insert into
 537  * @insert_keys:        list of keys to insert
 538  * @hook:               insert callback
 539  */
 540 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
 541                      struct bkey_i *k,
 542                      struct disk_reservation *disk_res,
 543                      struct extent_insert_hook *hook,
 544                      u64 *journal_seq, int flags)
 545 {
 546         struct btree_iter iter;
 547         int ret;
 548
 549         bch2_btree_iter_init(&iter, c, id, bkey_start_pos(&k->k),
 550                              BTREE_ITER_INTENT);
 551         ret = bch2_btree_insert_at(c, disk_res, hook, journal_seq, flags,
 552                                   BTREE_INSERT_ENTRY(&iter, k));
 553         bch2_btree_iter_unlock(&iter);
 554
 555         return ret;
 556 }
 557
 558 /*
 559  * bch_btree_delete_range - delete everything within a given range
 560  *
 561  * Range is a half open interval - [start, end)
 562  */
 563 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 564                            struct bpos start,
 565                            struct bpos end,
 566                            struct bversion version,
 567                            struct disk_reservation *disk_res,
 568                            struct extent_insert_hook *hook,
 569                            u64 *journal_seq)
 570 {
 571         struct btree_iter iter;
 572         struct bkey_s_c k;
 573         int ret = 0;
 574
 575         bch2_btree_iter_init(&iter, c, id, start,
 576                              BTREE_ITER_INTENT);
 577
 578         while ((k = bch2_btree_iter_peek(&iter)).k &&
 579                !(ret = btree_iter_err(k))) {
 580                 unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits);
 581                 /* really shouldn't be using a bare, unpadded bkey_i */
 582                 struct bkey_i delete;
 583
 584                 if (bkey_cmp(iter.pos, end) >= 0)
 585                         break;
 586
 587                 bkey_init(&delete.k);
 588
 589                 /*
 590                  * For extents, iter.pos won't necessarily be the same as
 591                  * bkey_start_pos(k.k) (for non extents they always will be the
 592                  * same). It's important that we delete starting from iter.pos
 593                  * because the range we want to delete could start in the middle
 594                  * of k.
 595                  *
 596                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 597                  * bkey_start_pos(k.k)).
 598                  */
 599                 delete.k.p = iter.pos;
 600                 delete.k.version = version;
 601
 602                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
 603                         /*
 604                          * The extents btree is special - KEY_TYPE_DISCARD is
 605                          * used for deletions, not KEY_TYPE_DELETED. This is an
 606                          * internal implementation detail that probably
 607                          * shouldn't be exposed (internally, KEY_TYPE_DELETED is
 608                          * used as a proxy for k->size == 0):
 609                          */
 610                         delete.k.type = KEY_TYPE_DISCARD;
 611
 612                         /* create the biggest key we can */
 613                         bch2_key_resize(&delete.k, max_sectors);
 614                         bch2_cut_back(end, &delete.k);
 615                 }
 616
 617                 ret = bch2_btree_insert_at(c, disk_res, hook, journal_seq,
 618                                           BTREE_INSERT_NOFAIL,
 619                                           BTREE_INSERT_ENTRY(&iter, &delete));
 620                 if (ret)
 621                         break;
 622
 623                 bch2_btree_iter_cond_resched(&iter);
 624         }
 625
 626         bch2_btree_iter_unlock(&iter);
 627         return ret;
 628 }