fs/bcachefs/bset.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Code for working with individual keys, and sorted sets of keys with in a
   4  * btree node
   5  *
   6  * Copyright 2012 Google, Inc.
   7  */
   8
   9 #include "bcachefs.h"
  10 #include "btree_cache.h"
  11 #include "bset.h"
  12 #include "eytzinger.h"
  13 #include "trace.h"
  14 #include "util.h"
  15
  16 #include <linux/unaligned.h>
  17 #include <linux/console.h>
  18 #include <linux/random.h>
  19 #include <linux/prefetch.h>
  20
  21 static inline void __bch2_btree_node_iter_advance(struct btree_node_iter *,
  22                                                   struct btree *);
  23
  24 static inline unsigned __btree_node_iter_used(struct btree_node_iter *iter)
  25 {
  26         unsigned n = ARRAY_SIZE(iter->data);
  27
  28         while (n && __btree_node_iter_set_end(iter, n - 1))
  29                 --n;
  30
  31         return n;
  32 }
  33
  34 struct bset_tree *bch2_bkey_to_bset(struct btree *b, struct bkey_packed *k)
  35 {
  36         return bch2_bkey_to_bset_inlined(b, k);
  37 }
  38
  39 /*
  40  * There are never duplicate live keys in the btree - but including keys that
  41  * have been flagged as deleted (and will be cleaned up later) we _will_ see
  42  * duplicates.
  43  *
  44  * Thus the sort order is: usual key comparison first, but for keys that compare
  45  * equal the deleted key(s) come first, and the (at most one) live version comes
  46  * last.
  47  *
  48  * The main reason for this is insertion: to handle overwrites, we first iterate
  49  * over keys that compare equal to our insert key, and then insert immediately
  50  * prior to the first key greater than the key we're inserting - our insert
  51  * position will be after all keys that compare equal to our insert key, which
  52  * by the time we actually do the insert will all be deleted.
  53  */
  54
  55 void bch2_dump_bset(struct bch_fs *c, struct btree *b,
  56                     struct bset *i, unsigned set)
  57 {
  58         struct bkey_packed *_k, *_n;
  59         struct bkey uk, n;
  60         struct bkey_s_c k;
  61         struct printbuf buf = PRINTBUF;
  62
  63         if (!i->u64s)
  64                 return;
  65
  66         for (_k = i->start;
  67              _k < vstruct_last(i);
  68              _k = _n) {
  69                 _n = bkey_p_next(_k);
  70
  71                 if (!_k->u64s) {
  72                         printk(KERN_ERR "block %u key %5zu - u64s 0? aieee!\n", set,
  73                                _k->_data - i->_data);
  74                         break;
  75                 }
  76
  77                 k = bkey_disassemble(b, _k, &uk);
  78
  79                 printbuf_reset(&buf);
  80                 if (c)
  81                         bch2_bkey_val_to_text(&buf, c, k);
  82                 else
  83                         bch2_bkey_to_text(&buf, k.k);
  84                 printk(KERN_ERR "block %u key %5zu: %s\n", set,
  85                        _k->_data - i->_data, buf.buf);
  86
  87                 if (_n == vstruct_last(i))
  88                         continue;
  89
  90                 n = bkey_unpack_key(b, _n);
  91
  92                 if (bpos_lt(n.p, k.k->p)) {
  93                         printk(KERN_ERR "Key skipped backwards\n");
  94                         continue;
  95                 }
  96
  97                 if (!bkey_deleted(k.k) && bpos_eq(n.p, k.k->p))
  98                         printk(KERN_ERR "Duplicate keys\n");
  99         }
 100
 101         printbuf_exit(&buf);
 102 }
 103
 104 void bch2_dump_btree_node(struct bch_fs *c, struct btree *b)
 105 {
 106         console_lock();
 107         for_each_bset(b, t)
 108                 bch2_dump_bset(c, b, bset(b, t), t - b->set);
 109         console_unlock();
 110 }
 111
 112 void bch2_dump_btree_node_iter(struct btree *b,
 113                               struct btree_node_iter *iter)
 114 {
 115         struct btree_node_iter_set *set;
 116         struct printbuf buf = PRINTBUF;
 117
 118         printk(KERN_ERR "btree node iter with %u/%u sets:\n",
 119                __btree_node_iter_used(iter), b->nsets);
 120
 121         btree_node_iter_for_each(iter, set) {
 122                 struct bkey_packed *k = __btree_node_offset_to_key(b, set->k);
 123                 struct bset_tree *t = bch2_bkey_to_bset(b, k);
 124                 struct bkey uk = bkey_unpack_key(b, k);
 125
 126                 printbuf_reset(&buf);
 127                 bch2_bkey_to_text(&buf, &uk);
 128                 printk(KERN_ERR "set %zu key %u: %s\n",
 129                        t - b->set, set->k, buf.buf);
 130         }
 131
 132         printbuf_exit(&buf);
 133 }
 134
 135 struct btree_nr_keys bch2_btree_node_count_keys(struct btree *b)
 136 {
 137         struct bkey_packed *k;
 138         struct btree_nr_keys nr = {};
 139
 140         for_each_bset(b, t)
 141                 bset_tree_for_each_key(b, t, k)
 142                         if (!bkey_deleted(k))
 143                                 btree_keys_account_key_add(&nr, t - b->set, k);
 144         return nr;
 145 }
 146
 147 #ifdef CONFIG_BCACHEFS_DEBUG
 148
 149 void __bch2_verify_btree_nr_keys(struct btree *b)
 150 {
 151         struct btree_nr_keys nr = bch2_btree_node_count_keys(b);
 152
 153         BUG_ON(memcmp(&nr, &b->nr, sizeof(nr)));
 154 }
 155
 156 static void bch2_btree_node_iter_next_check(struct btree_node_iter *_iter,
 157                                             struct btree *b)
 158 {
 159         struct btree_node_iter iter = *_iter;
 160         const struct bkey_packed *k, *n;
 161
 162         k = bch2_btree_node_iter_peek_all(&iter, b);
 163         __bch2_btree_node_iter_advance(&iter, b);
 164         n = bch2_btree_node_iter_peek_all(&iter, b);
 165
 166         bkey_unpack_key(b, k);
 167
 168         if (n &&
 169             bkey_iter_cmp(b, k, n) > 0) {
 170                 struct btree_node_iter_set *set;
 171                 struct bkey ku = bkey_unpack_key(b, k);
 172                 struct bkey nu = bkey_unpack_key(b, n);
 173                 struct printbuf buf1 = PRINTBUF;
 174                 struct printbuf buf2 = PRINTBUF;
 175
 176                 bch2_dump_btree_node(NULL, b);
 177                 bch2_bkey_to_text(&buf1, &ku);
 178                 bch2_bkey_to_text(&buf2, &nu);
 179                 printk(KERN_ERR "out of order/overlapping:\n%s\n%s\n",
 180                        buf1.buf, buf2.buf);
 181                 printk(KERN_ERR "iter was:");
 182
 183                 btree_node_iter_for_each(_iter, set) {
 184                         struct bkey_packed *k2 = __btree_node_offset_to_key(b, set->k);
 185                         struct bset_tree *t = bch2_bkey_to_bset(b, k2);
 186                         printk(" [%zi %zi]", t - b->set,
 187                                k2->_data - bset(b, t)->_data);
 188                 }
 189                 panic("\n");
 190         }
 191 }
 192
 193 void bch2_btree_node_iter_verify(struct btree_node_iter *iter,
 194                                  struct btree *b)
 195 {
 196         struct btree_node_iter_set *set, *s2;
 197         struct bkey_packed *k, *p;
 198
 199         if (bch2_btree_node_iter_end(iter))
 200                 return;
 201
 202         /* Verify no duplicates: */
 203         btree_node_iter_for_each(iter, set) {
 204                 BUG_ON(set->k > set->end);
 205                 btree_node_iter_for_each(iter, s2)
 206                         BUG_ON(set != s2 && set->end == s2->end);
 207         }
 208
 209         /* Verify that set->end is correct: */
 210         btree_node_iter_for_each(iter, set) {
 211                 for_each_bset(b, t)
 212                         if (set->end == t->end_offset) {
 213                                 BUG_ON(set->k < btree_bkey_first_offset(t) ||
 214                                        set->k >= t->end_offset);
 215                                 goto found;
 216                         }
 217                 BUG();
 218 found:
 219                 do {} while (0);
 220         }
 221
 222         /* Verify iterator is sorted: */
 223         btree_node_iter_for_each(iter, set)
 224                 BUG_ON(set != iter->data &&
 225                        btree_node_iter_cmp(b, set[-1], set[0]) > 0);
 226
 227         k = bch2_btree_node_iter_peek_all(iter, b);
 228
 229         for_each_bset(b, t) {
 230                 if (iter->data[0].end == t->end_offset)
 231                         continue;
 232
 233                 p = bch2_bkey_prev_all(b, t,
 234                         bch2_btree_node_iter_bset_pos(iter, b, t));
 235
 236                 BUG_ON(p && bkey_iter_cmp(b, k, p) < 0);
 237         }
 238 }
 239
 240 void bch2_verify_insert_pos(struct btree *b, struct bkey_packed *where,
 241                             struct bkey_packed *insert, unsigned clobber_u64s)
 242 {
 243         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 244         struct bkey_packed *prev = bch2_bkey_prev_all(b, t, where);
 245         struct bkey_packed *next = (void *) ((u64 *) where->_data + clobber_u64s);
 246         struct printbuf buf1 = PRINTBUF;
 247         struct printbuf buf2 = PRINTBUF;
 248 #if 0
 249         BUG_ON(prev &&
 250                bkey_iter_cmp(b, prev, insert) > 0);
 251 #else
 252         if (prev &&
 253             bkey_iter_cmp(b, prev, insert) > 0) {
 254                 struct bkey k1 = bkey_unpack_key(b, prev);
 255                 struct bkey k2 = bkey_unpack_key(b, insert);
 256
 257                 bch2_dump_btree_node(NULL, b);
 258                 bch2_bkey_to_text(&buf1, &k1);
 259                 bch2_bkey_to_text(&buf2, &k2);
 260
 261                 panic("prev > insert:\n"
 262                       "prev    key %s\n"
 263                       "insert  key %s\n",
 264                       buf1.buf, buf2.buf);
 265         }
 266 #endif
 267 #if 0
 268         BUG_ON(next != btree_bkey_last(b, t) &&
 269                bkey_iter_cmp(b, insert, next) > 0);
 270 #else
 271         if (next != btree_bkey_last(b, t) &&
 272             bkey_iter_cmp(b, insert, next) > 0) {
 273                 struct bkey k1 = bkey_unpack_key(b, insert);
 274                 struct bkey k2 = bkey_unpack_key(b, next);
 275
 276                 bch2_dump_btree_node(NULL, b);
 277                 bch2_bkey_to_text(&buf1, &k1);
 278                 bch2_bkey_to_text(&buf2, &k2);
 279
 280                 panic("insert > next:\n"
 281                       "insert  key %s\n"
 282                       "next    key %s\n",
 283                       buf1.buf, buf2.buf);
 284         }
 285 #endif
 286 }
 287
 288 #else
 289
 290 static inline void bch2_btree_node_iter_next_check(struct btree_node_iter *iter,
 291                                                    struct btree *b) {}
 292
 293 #endif
 294
 295 /* Auxiliary search trees */
 296
 297 #define BFLOAT_FAILED_UNPACKED  U8_MAX
 298 #define BFLOAT_FAILED           U8_MAX
 299
 300 struct bkey_float {
 301         u8              exponent;
 302         u8              key_offset;
 303         u16             mantissa;
 304 };
 305 #define BKEY_MANTISSA_BITS      16
 306
 307 struct ro_aux_tree {
 308         u8                      nothing[0];
 309         struct bkey_float       f[];
 310 };
 311
 312 struct rw_aux_tree {
 313         u16             offset;
 314         struct bpos     k;
 315 };
 316
 317 static unsigned bset_aux_tree_buf_end(const struct bset_tree *t)
 318 {
 319         BUG_ON(t->aux_data_offset == U16_MAX);
 320
 321         switch (bset_aux_tree_type(t)) {
 322         case BSET_NO_AUX_TREE:
 323                 return t->aux_data_offset;
 324         case BSET_RO_AUX_TREE:
 325                 return t->aux_data_offset +
 326                         DIV_ROUND_UP(t->size * sizeof(struct bkey_float), 8);
 327         case BSET_RW_AUX_TREE:
 328                 return t->aux_data_offset +
 329                         DIV_ROUND_UP(sizeof(struct rw_aux_tree) * t->size, 8);
 330         default:
 331                 BUG();
 332         }
 333 }
 334
 335 static unsigned bset_aux_tree_buf_start(const struct btree *b,
 336                                         const struct bset_tree *t)
 337 {
 338         return t == b->set
 339                 ? DIV_ROUND_UP(b->unpack_fn_len, 8)
 340                 : bset_aux_tree_buf_end(t - 1);
 341 }
 342
 343 static void *__aux_tree_base(const struct btree *b,
 344                              const struct bset_tree *t)
 345 {
 346         return b->aux_data + t->aux_data_offset * 8;
 347 }
 348
 349 static struct ro_aux_tree *ro_aux_tree_base(const struct btree *b,
 350                                             const struct bset_tree *t)
 351 {
 352         EBUG_ON(bset_aux_tree_type(t) != BSET_RO_AUX_TREE);
 353
 354         return __aux_tree_base(b, t);
 355 }
 356
 357 static struct bkey_float *bkey_float(const struct btree *b,
 358                                      const struct bset_tree *t,
 359                                      unsigned idx)
 360 {
 361         return ro_aux_tree_base(b, t)->f + idx;
 362 }
 363
 364 static void bset_aux_tree_verify(struct btree *b)
 365 {
 366 #ifdef CONFIG_BCACHEFS_DEBUG
 367         for_each_bset(b, t) {
 368                 if (t->aux_data_offset == U16_MAX)
 369                         continue;
 370
 371                 BUG_ON(t != b->set &&
 372                        t[-1].aux_data_offset == U16_MAX);
 373
 374                 BUG_ON(t->aux_data_offset < bset_aux_tree_buf_start(b, t));
 375                 BUG_ON(t->aux_data_offset > btree_aux_data_u64s(b));
 376                 BUG_ON(bset_aux_tree_buf_end(t) > btree_aux_data_u64s(b));
 377         }
 378 #endif
 379 }
 380
 381 void bch2_btree_keys_init(struct btree *b)
 382 {
 383         unsigned i;
 384
 385         b->nsets                = 0;
 386         memset(&b->nr, 0, sizeof(b->nr));
 387
 388         for (i = 0; i < MAX_BSETS; i++)
 389                 b->set[i].data_offset = U16_MAX;
 390
 391         bch2_bset_set_no_aux_tree(b, b->set);
 392 }
 393
 394 /* Binary tree stuff for auxiliary search trees */
 395
 396 /*
 397  * Cacheline/offset <-> bkey pointer arithmetic:
 398  *
 399  * t->tree is a binary search tree in an array; each node corresponds to a key
 400  * in one cacheline in t->set (BSET_CACHELINE bytes).
 401  *
 402  * This means we don't have to store the full index of the key that a node in
 403  * the binary tree points to; eytzinger1_to_inorder() gives us the cacheline, and
 404  * then bkey_float->m gives us the offset within that cacheline, in units of 8
 405  * bytes.
 406  *
 407  * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
 408  * make this work.
 409  *
 410  * To construct the bfloat for an arbitrary key we need to know what the key
 411  * immediately preceding it is: we have to check if the two keys differ in the
 412  * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
 413  * of the previous key so we can walk backwards to it from t->tree[j]'s key.
 414  */
 415
 416 static inline void *bset_cacheline(const struct btree *b,
 417                                    const struct bset_tree *t,
 418                                    unsigned cacheline)
 419 {
 420         return (void *) round_down((unsigned long) btree_bkey_first(b, t),
 421                                    L1_CACHE_BYTES) +
 422                 cacheline * BSET_CACHELINE;
 423 }
 424
 425 static struct bkey_packed *cacheline_to_bkey(const struct btree *b,
 426                                              const struct bset_tree *t,
 427                                              unsigned cacheline,
 428                                              unsigned offset)
 429 {
 430         return bset_cacheline(b, t, cacheline) + offset * 8;
 431 }
 432
 433 static unsigned bkey_to_cacheline(const struct btree *b,
 434                                   const struct bset_tree *t,
 435                                   const struct bkey_packed *k)
 436 {
 437         return ((void *) k - bset_cacheline(b, t, 0)) / BSET_CACHELINE;
 438 }
 439
 440 static ssize_t __bkey_to_cacheline_offset(const struct btree *b,
 441                                           const struct bset_tree *t,
 442                                           unsigned cacheline,
 443                                           const struct bkey_packed *k)
 444 {
 445         return (u64 *) k - (u64 *) bset_cacheline(b, t, cacheline);
 446 }
 447
 448 static unsigned bkey_to_cacheline_offset(const struct btree *b,
 449                                          const struct bset_tree *t,
 450                                          unsigned cacheline,
 451                                          const struct bkey_packed *k)
 452 {
 453         size_t m = __bkey_to_cacheline_offset(b, t, cacheline, k);
 454
 455         EBUG_ON(m > U8_MAX);
 456         return m;
 457 }
 458
 459 static inline struct bkey_packed *tree_to_bkey(const struct btree *b,
 460                                                const struct bset_tree *t,
 461                                                unsigned j)
 462 {
 463         return cacheline_to_bkey(b, t,
 464                         __eytzinger1_to_inorder(j, t->size - 1, t->extra),
 465                         bkey_float(b, t, j)->key_offset);
 466 }
 467
 468 static struct rw_aux_tree *rw_aux_tree(const struct btree *b,
 469                                        const struct bset_tree *t)
 470 {
 471         EBUG_ON(bset_aux_tree_type(t) != BSET_RW_AUX_TREE);
 472
 473         return __aux_tree_base(b, t);
 474 }
 475
 476 /*
 477  * For the write set - the one we're currently inserting keys into - we don't
 478  * maintain a full search tree, we just keep a simple lookup table in t->prev.
 479  */
 480 static struct bkey_packed *rw_aux_to_bkey(const struct btree *b,
 481                                           struct bset_tree *t,
 482                                           unsigned j)
 483 {
 484         return __btree_node_offset_to_key(b, rw_aux_tree(b, t)[j].offset);
 485 }
 486
 487 static void rw_aux_tree_set(const struct btree *b, struct bset_tree *t,
 488                             unsigned j, struct bkey_packed *k)
 489 {
 490         EBUG_ON(k >= btree_bkey_last(b, t));
 491
 492         rw_aux_tree(b, t)[j] = (struct rw_aux_tree) {
 493                 .offset = __btree_node_key_to_offset(b, k),
 494                 .k      = bkey_unpack_pos(b, k),
 495         };
 496 }
 497
 498 static void bch2_bset_verify_rw_aux_tree(struct btree *b,
 499                                         struct bset_tree *t)
 500 {
 501         struct bkey_packed *k = btree_bkey_first(b, t);
 502         unsigned j = 0;
 503
 504         if (!bch2_expensive_debug_checks)
 505                 return;
 506
 507         BUG_ON(bset_has_ro_aux_tree(t));
 508
 509         if (!bset_has_rw_aux_tree(t))
 510                 return;
 511
 512         BUG_ON(t->size < 1);
 513         BUG_ON(rw_aux_to_bkey(b, t, j) != k);
 514
 515         goto start;
 516         while (1) {
 517                 if (rw_aux_to_bkey(b, t, j) == k) {
 518                         BUG_ON(!bpos_eq(rw_aux_tree(b, t)[j].k,
 519                                         bkey_unpack_pos(b, k)));
 520 start:
 521                         if (++j == t->size)
 522                                 break;
 523
 524                         BUG_ON(rw_aux_tree(b, t)[j].offset <=
 525                                rw_aux_tree(b, t)[j - 1].offset);
 526                 }
 527
 528                 k = bkey_p_next(k);
 529                 BUG_ON(k >= btree_bkey_last(b, t));
 530         }
 531 }
 532
 533 /* returns idx of first entry >= offset: */
 534 static unsigned rw_aux_tree_bsearch(struct btree *b,
 535                                     struct bset_tree *t,
 536                                     unsigned offset)
 537 {
 538         unsigned bset_offs = offset - btree_bkey_first_offset(t);
 539         unsigned bset_u64s = t->end_offset - btree_bkey_first_offset(t);
 540         unsigned idx = bset_u64s ? bset_offs * t->size / bset_u64s : 0;
 541
 542         EBUG_ON(bset_aux_tree_type(t) != BSET_RW_AUX_TREE);
 543         EBUG_ON(!t->size);
 544         EBUG_ON(idx > t->size);
 545
 546         while (idx < t->size &&
 547                rw_aux_tree(b, t)[idx].offset < offset)
 548                 idx++;
 549
 550         while (idx &&
 551                rw_aux_tree(b, t)[idx - 1].offset >= offset)
 552                 idx--;
 553
 554         EBUG_ON(idx < t->size &&
 555                 rw_aux_tree(b, t)[idx].offset < offset);
 556         EBUG_ON(idx && rw_aux_tree(b, t)[idx - 1].offset >= offset);
 557         EBUG_ON(idx + 1 < t->size &&
 558                 rw_aux_tree(b, t)[idx].offset ==
 559                 rw_aux_tree(b, t)[idx + 1].offset);
 560
 561         return idx;
 562 }
 563
 564 static inline unsigned bkey_mantissa(const struct bkey_packed *k,
 565                                      const struct bkey_float *f)
 566 {
 567         u64 v;
 568
 569         EBUG_ON(!bkey_packed(k));
 570
 571         v = get_unaligned((u64 *) (((u8 *) k->_data) + (f->exponent >> 3)));
 572
 573         /*
 574          * In little endian, we're shifting off low bits (and then the bits we
 575          * want are at the low end), in big endian we're shifting off high bits
 576          * (and then the bits we want are at the high end, so we shift them
 577          * back down):
 578          */
 579 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 580         v >>= f->exponent & 7;
 581 #else
 582         v >>= 64 - (f->exponent & 7) - BKEY_MANTISSA_BITS;
 583 #endif
 584         return (u16) v;
 585 }
 586
 587 static __always_inline void make_bfloat(struct btree *b, struct bset_tree *t,
 588                                         unsigned j,
 589                                         struct bkey_packed *min_key,
 590                                         struct bkey_packed *max_key)
 591 {
 592         struct bkey_float *f = bkey_float(b, t, j);
 593         struct bkey_packed *m = tree_to_bkey(b, t, j);
 594         struct bkey_packed *l = is_power_of_2(j)
 595                 ? min_key
 596                 : tree_to_bkey(b, t, j >> ffs(j));
 597         struct bkey_packed *r = is_power_of_2(j + 1)
 598                 ? max_key
 599                 : tree_to_bkey(b, t, j >> (ffz(j) + 1));
 600         unsigned mantissa;
 601         int shift, exponent, high_bit;
 602
 603         /*
 604          * for failed bfloats, the lookup code falls back to comparing against
 605          * the original key.
 606          */
 607
 608         if (!bkey_packed(l) || !bkey_packed(r) || !bkey_packed(m) ||
 609             !b->nr_key_bits) {
 610                 f->exponent = BFLOAT_FAILED_UNPACKED;
 611                 return;
 612         }
 613
 614         /*
 615          * The greatest differing bit of l and r is the first bit we must
 616          * include in the bfloat mantissa we're creating in order to do
 617          * comparisons - that bit always becomes the high bit of
 618          * bfloat->mantissa, and thus the exponent we're calculating here is
 619          * the position of what will become the low bit in bfloat->mantissa:
 620          *
 621          * Note that this may be negative - we may be running off the low end
 622          * of the key: we handle this later:
 623          */
 624         high_bit = max(bch2_bkey_greatest_differing_bit(b, l, r),
 625                        min_t(unsigned, BKEY_MANTISSA_BITS, b->nr_key_bits) - 1);
 626         exponent = high_bit - (BKEY_MANTISSA_BITS - 1);
 627
 628         /*
 629          * Then we calculate the actual shift value, from the start of the key
 630          * (k->_data), to get the key bits starting at exponent:
 631          */
 632 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 633         shift = (int) (b->format.key_u64s * 64 - b->nr_key_bits) + exponent;
 634
 635         EBUG_ON(shift + BKEY_MANTISSA_BITS > b->format.key_u64s * 64);
 636 #else
 637         shift = high_bit_offset +
 638                 b->nr_key_bits -
 639                 exponent -
 640                 BKEY_MANTISSA_BITS;
 641
 642         EBUG_ON(shift < KEY_PACKED_BITS_START);
 643 #endif
 644         EBUG_ON(shift < 0 || shift >= BFLOAT_FAILED);
 645
 646         f->exponent = shift;
 647         mantissa = bkey_mantissa(m, f);
 648
 649         /*
 650          * If we've got garbage bits, set them to all 1s - it's legal for the
 651          * bfloat to compare larger than the original key, but not smaller:
 652          */
 653         if (exponent < 0)
 654                 mantissa |= ~(~0U << -exponent);
 655
 656         f->mantissa = mantissa;
 657 }
 658
 659 /* bytes remaining - only valid for last bset: */
 660 static unsigned __bset_tree_capacity(struct btree *b, const struct bset_tree *t)
 661 {
 662         bset_aux_tree_verify(b);
 663
 664         return btree_aux_data_bytes(b) - t->aux_data_offset * sizeof(u64);
 665 }
 666
 667 static unsigned bset_ro_tree_capacity(struct btree *b, const struct bset_tree *t)
 668 {
 669         return __bset_tree_capacity(b, t) / sizeof(struct bkey_float);
 670 }
 671
 672 static unsigned bset_rw_tree_capacity(struct btree *b, const struct bset_tree *t)
 673 {
 674         return __bset_tree_capacity(b, t) / sizeof(struct rw_aux_tree);
 675 }
 676
 677 static noinline void __build_rw_aux_tree(struct btree *b, struct bset_tree *t)
 678 {
 679         struct bkey_packed *k;
 680
 681         t->size = 1;
 682         t->extra = BSET_RW_AUX_TREE_VAL;
 683         rw_aux_tree(b, t)[0].offset =
 684                 __btree_node_key_to_offset(b, btree_bkey_first(b, t));
 685
 686         bset_tree_for_each_key(b, t, k) {
 687                 if (t->size == bset_rw_tree_capacity(b, t))
 688                         break;
 689
 690                 if ((void *) k - (void *) rw_aux_to_bkey(b, t, t->size - 1) >
 691                     L1_CACHE_BYTES)
 692                         rw_aux_tree_set(b, t, t->size++, k);
 693         }
 694 }
 695
 696 static noinline void __build_ro_aux_tree(struct btree *b, struct bset_tree *t)
 697 {
 698         struct bkey_packed *k = btree_bkey_first(b, t);
 699         struct bkey_i min_key, max_key;
 700         unsigned cacheline = 1;
 701
 702         t->size = min(bkey_to_cacheline(b, t, btree_bkey_last(b, t)),
 703                       bset_ro_tree_capacity(b, t));
 704 retry:
 705         if (t->size < 2) {
 706                 t->size = 0;
 707                 t->extra = BSET_NO_AUX_TREE_VAL;
 708                 return;
 709         }
 710
 711         t->extra = eytzinger1_extra(t->size - 1);
 712
 713         /* First we figure out where the first key in each cacheline is */
 714         eytzinger1_for_each(j, t->size - 1) {
 715                 while (bkey_to_cacheline(b, t, k) < cacheline)
 716                         k = bkey_p_next(k);
 717
 718                 if (k >= btree_bkey_last(b, t)) {
 719                         /* XXX: this path sucks */
 720                         t->size--;
 721                         goto retry;
 722                 }
 723
 724                 bkey_float(b, t, j)->key_offset =
 725                         bkey_to_cacheline_offset(b, t, cacheline++, k);
 726
 727                 EBUG_ON(tree_to_bkey(b, t, j) != k);
 728         }
 729
 730         if (!bkey_pack_pos(bkey_to_packed(&min_key), b->data->min_key, b)) {
 731                 bkey_init(&min_key.k);
 732                 min_key.k.p = b->data->min_key;
 733         }
 734
 735         if (!bkey_pack_pos(bkey_to_packed(&max_key), b->data->max_key, b)) {
 736                 bkey_init(&max_key.k);
 737                 max_key.k.p = b->data->max_key;
 738         }
 739
 740         /* Then we build the tree */
 741         eytzinger1_for_each(j, t->size - 1)
 742                 make_bfloat(b, t, j,
 743                             bkey_to_packed(&min_key),
 744                             bkey_to_packed(&max_key));
 745 }
 746
 747 static void bset_alloc_tree(struct btree *b, struct bset_tree *t)
 748 {
 749         struct bset_tree *i;
 750
 751         for (i = b->set; i != t; i++)
 752                 BUG_ON(bset_has_rw_aux_tree(i));
 753
 754         bch2_bset_set_no_aux_tree(b, t);
 755
 756         /* round up to next cacheline: */
 757         t->aux_data_offset = round_up(bset_aux_tree_buf_start(b, t),
 758                                       SMP_CACHE_BYTES / sizeof(u64));
 759
 760         bset_aux_tree_verify(b);
 761 }
 762
 763 void bch2_bset_build_aux_tree(struct btree *b, struct bset_tree *t,
 764                              bool writeable)
 765 {
 766         if (writeable
 767             ? bset_has_rw_aux_tree(t)
 768             : bset_has_ro_aux_tree(t))
 769                 return;
 770
 771         bset_alloc_tree(b, t);
 772
 773         if (!__bset_tree_capacity(b, t))
 774                 return;
 775
 776         if (writeable)
 777                 __build_rw_aux_tree(b, t);
 778         else
 779                 __build_ro_aux_tree(b, t);
 780
 781         bset_aux_tree_verify(b);
 782 }
 783
 784 void bch2_bset_init_first(struct btree *b, struct bset *i)
 785 {
 786         struct bset_tree *t;
 787
 788         BUG_ON(b->nsets);
 789
 790         memset(i, 0, sizeof(*i));
 791         get_random_bytes(&i->seq, sizeof(i->seq));
 792         SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 793
 794         t = &b->set[b->nsets++];
 795         set_btree_bset(b, t, i);
 796 }
 797
 798 void bch2_bset_init_next(struct btree *b, struct btree_node_entry *bne)
 799 {
 800         struct bset *i = &bne->keys;
 801         struct bset_tree *t;
 802
 803         BUG_ON(bset_byte_offset(b, bne) >= btree_buf_bytes(b));
 804         BUG_ON((void *) bne < (void *) btree_bkey_last(b, bset_tree_last(b)));
 805         BUG_ON(b->nsets >= MAX_BSETS);
 806
 807         memset(i, 0, sizeof(*i));
 808         i->seq = btree_bset_first(b)->seq;
 809         SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 810
 811         t = &b->set[b->nsets++];
 812         set_btree_bset(b, t, i);
 813 }
 814
 815 /*
 816  * find _some_ key in the same bset as @k that precedes @k - not necessarily the
 817  * immediate predecessor:
 818  */
 819 static struct bkey_packed *__bkey_prev(struct btree *b, struct bset_tree *t,
 820                                        struct bkey_packed *k)
 821 {
 822         struct bkey_packed *p;
 823         unsigned offset;
 824         int j;
 825
 826         EBUG_ON(k < btree_bkey_first(b, t) ||
 827                 k > btree_bkey_last(b, t));
 828
 829         if (k == btree_bkey_first(b, t))
 830                 return NULL;
 831
 832         switch (bset_aux_tree_type(t)) {
 833         case BSET_NO_AUX_TREE:
 834                 p = btree_bkey_first(b, t);
 835                 break;
 836         case BSET_RO_AUX_TREE:
 837                 j = min_t(unsigned, t->size - 1, bkey_to_cacheline(b, t, k));
 838
 839                 do {
 840                         p = j ? tree_to_bkey(b, t,
 841                                         __inorder_to_eytzinger1(j--,
 842                                                         t->size - 1, t->extra))
 843                               : btree_bkey_first(b, t);
 844                 } while (p >= k);
 845                 break;
 846         case BSET_RW_AUX_TREE:
 847                 offset = __btree_node_key_to_offset(b, k);
 848                 j = rw_aux_tree_bsearch(b, t, offset);
 849                 p = j ? rw_aux_to_bkey(b, t, j - 1)
 850                       : btree_bkey_first(b, t);
 851                 break;
 852         }
 853
 854         return p;
 855 }
 856
 857 struct bkey_packed *bch2_bkey_prev_filter(struct btree *b,
 858                                           struct bset_tree *t,
 859                                           struct bkey_packed *k,
 860                                           unsigned min_key_type)
 861 {
 862         struct bkey_packed *p, *i, *ret = NULL, *orig_k = k;
 863
 864         while ((p = __bkey_prev(b, t, k)) && !ret) {
 865                 for (i = p; i != k; i = bkey_p_next(i))
 866                         if (i->type >= min_key_type)
 867                                 ret = i;
 868
 869                 k = p;
 870         }
 871
 872         if (bch2_expensive_debug_checks) {
 873                 BUG_ON(ret >= orig_k);
 874
 875                 for (i = ret
 876                         ? bkey_p_next(ret)
 877                         : btree_bkey_first(b, t);
 878                      i != orig_k;
 879                      i = bkey_p_next(i))
 880                         BUG_ON(i->type >= min_key_type);
 881         }
 882
 883         return ret;
 884 }
 885
 886 /* Insert */
 887
 888 static void rw_aux_tree_insert_entry(struct btree *b,
 889                                      struct bset_tree *t,
 890                                      unsigned idx)
 891 {
 892         EBUG_ON(!idx || idx > t->size);
 893         struct bkey_packed *start = rw_aux_to_bkey(b, t, idx - 1);
 894         struct bkey_packed *end = idx < t->size
 895                                   ? rw_aux_to_bkey(b, t, idx)
 896                                   : btree_bkey_last(b, t);
 897
 898         if (t->size < bset_rw_tree_capacity(b, t) &&
 899             (void *) end - (void *) start > L1_CACHE_BYTES) {
 900                 struct bkey_packed *k = start;
 901
 902                 while (1) {
 903                         k = bkey_p_next(k);
 904                         if (k == end)
 905                                 break;
 906
 907                         if ((void *) k - (void *) start >= L1_CACHE_BYTES) {
 908                                 memmove(&rw_aux_tree(b, t)[idx + 1],
 909                                         &rw_aux_tree(b, t)[idx],
 910                                         (void *) &rw_aux_tree(b, t)[t->size] -
 911                                         (void *) &rw_aux_tree(b, t)[idx]);
 912                                 t->size++;
 913                                 rw_aux_tree_set(b, t, idx, k);
 914                                 break;
 915                         }
 916                 }
 917         }
 918 }
 919
 920 static void bch2_bset_fix_lookup_table(struct btree *b,
 921                                        struct bset_tree *t,
 922                                        struct bkey_packed *_where,
 923                                        unsigned clobber_u64s,
 924                                        unsigned new_u64s)
 925 {
 926         int shift = new_u64s - clobber_u64s;
 927         unsigned idx, j, where = __btree_node_key_to_offset(b, _where);
 928
 929         EBUG_ON(bset_has_ro_aux_tree(t));
 930
 931         if (!bset_has_rw_aux_tree(t))
 932                 return;
 933
 934         if (where > rw_aux_tree(b, t)[t->size - 1].offset) {
 935                 rw_aux_tree_insert_entry(b, t, t->size);
 936                 goto verify;
 937         }
 938
 939         /* returns first entry >= where */
 940         idx = rw_aux_tree_bsearch(b, t, where);
 941
 942         if (rw_aux_tree(b, t)[idx].offset == where) {
 943                 if (!idx) { /* never delete first entry */
 944                         idx++;
 945                 } else if (where < t->end_offset) {
 946                         rw_aux_tree_set(b, t, idx++, _where);
 947                 } else {
 948                         EBUG_ON(where != t->end_offset);
 949                         rw_aux_tree_insert_entry(b, t, --t->size);
 950                         goto verify;
 951                 }
 952         }
 953
 954         EBUG_ON(idx < t->size && rw_aux_tree(b, t)[idx].offset <= where);
 955         if (idx < t->size &&
 956             rw_aux_tree(b, t)[idx].offset + shift ==
 957             rw_aux_tree(b, t)[idx - 1].offset) {
 958                 memmove(&rw_aux_tree(b, t)[idx],
 959                         &rw_aux_tree(b, t)[idx + 1],
 960                         (void *) &rw_aux_tree(b, t)[t->size] -
 961                         (void *) &rw_aux_tree(b, t)[idx + 1]);
 962                 t->size -= 1;
 963         }
 964
 965         for (j = idx; j < t->size; j++)
 966                 rw_aux_tree(b, t)[j].offset += shift;
 967
 968         EBUG_ON(idx < t->size &&
 969                 rw_aux_tree(b, t)[idx].offset ==
 970                 rw_aux_tree(b, t)[idx - 1].offset);
 971
 972         rw_aux_tree_insert_entry(b, t, idx);
 973
 974 verify:
 975         bch2_bset_verify_rw_aux_tree(b, t);
 976         bset_aux_tree_verify(b);
 977 }
 978
 979 void bch2_bset_insert(struct btree *b,
 980                       struct bkey_packed *where,
 981                       struct bkey_i *insert,
 982                       unsigned clobber_u64s)
 983 {
 984         struct bkey_format *f = &b->format;
 985         struct bset_tree *t = bset_tree_last(b);
 986         struct bkey_packed packed, *src = bkey_to_packed(insert);
 987
 988         bch2_bset_verify_rw_aux_tree(b, t);
 989         bch2_verify_insert_pos(b, where, bkey_to_packed(insert), clobber_u64s);
 990
 991         if (bch2_bkey_pack_key(&packed, &insert->k, f))
 992                 src = &packed;
 993
 994         if (!bkey_deleted(&insert->k))
 995                 btree_keys_account_key_add(&b->nr, t - b->set, src);
 996
 997         if (src->u64s != clobber_u64s) {
 998                 u64 *src_p = (u64 *) where->_data + clobber_u64s;
 999                 u64 *dst_p = (u64 *) where->_data + src->u64s;
1000
1001                 EBUG_ON((int) le16_to_cpu(bset(b, t)->u64s) <
1002                         (int) clobber_u64s - src->u64s);
1003
1004                 memmove_u64s(dst_p, src_p, btree_bkey_last(b, t)->_data - src_p);
1005                 le16_add_cpu(&bset(b, t)->u64s, src->u64s - clobber_u64s);
1006                 set_btree_bset_end(b, t);
1007         }
1008
1009         memcpy_u64s_small(where, src,
1010                     bkeyp_key_u64s(f, src));
1011         memcpy_u64s(bkeyp_val(f, where), &insert->v,
1012                     bkeyp_val_u64s(f, src));
1013
1014         if (src->u64s != clobber_u64s)
1015                 bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, src->u64s);
1016
1017         bch2_verify_btree_nr_keys(b);
1018 }
1019
1020 void bch2_bset_delete(struct btree *b,
1021                       struct bkey_packed *where,
1022                       unsigned clobber_u64s)
1023 {
1024         struct bset_tree *t = bset_tree_last(b);
1025         u64 *src_p = (u64 *) where->_data + clobber_u64s;
1026         u64 *dst_p = where->_data;
1027
1028         bch2_bset_verify_rw_aux_tree(b, t);
1029
1030         EBUG_ON(le16_to_cpu(bset(b, t)->u64s) < clobber_u64s);
1031
1032         memmove_u64s_down(dst_p, src_p, btree_bkey_last(b, t)->_data - src_p);
1033         le16_add_cpu(&bset(b, t)->u64s, -clobber_u64s);
1034         set_btree_bset_end(b, t);
1035
1036         bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, 0);
1037 }
1038
1039 /* Lookup */
1040
1041 __flatten
1042 static struct bkey_packed *bset_search_write_set(const struct btree *b,
1043                                 struct bset_tree *t,
1044                                 struct bpos *search)
1045 {
1046         unsigned l = 0, r = t->size;
1047
1048         while (l + 1 != r) {
1049                 unsigned m = (l + r) >> 1;
1050
1051                 if (bpos_lt(rw_aux_tree(b, t)[m].k, *search))
1052                         l = m;
1053                 else
1054                         r = m;
1055         }
1056
1057         return rw_aux_to_bkey(b, t, l);
1058 }
1059
1060 static inline void prefetch_four_cachelines(void *p)
1061 {
1062 #ifdef CONFIG_X86_64
1063         asm("prefetcht0 (-127 + 64 * 0)(%0);"
1064             "prefetcht0 (-127 + 64 * 1)(%0);"
1065             "prefetcht0 (-127 + 64 * 2)(%0);"
1066             "prefetcht0 (-127 + 64 * 3)(%0);"
1067             :
1068             : "r" (p + 127));
1069 #else
1070         prefetch(p + L1_CACHE_BYTES * 0);
1071         prefetch(p + L1_CACHE_BYTES * 1);
1072         prefetch(p + L1_CACHE_BYTES * 2);
1073         prefetch(p + L1_CACHE_BYTES * 3);
1074 #endif
1075 }
1076
1077 static inline bool bkey_mantissa_bits_dropped(const struct btree *b,
1078                                               const struct bkey_float *f)
1079 {
1080 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
1081         unsigned key_bits_start = b->format.key_u64s * 64 - b->nr_key_bits;
1082
1083         return f->exponent > key_bits_start;
1084 #else
1085         unsigned key_bits_end = high_bit_offset + b->nr_key_bits;
1086
1087         return f->exponent + BKEY_MANTISSA_BITS < key_bits_end;
1088 #endif
1089 }
1090
1091 __flatten
1092 static struct bkey_packed *bset_search_tree(const struct btree *b,
1093                                 const struct bset_tree *t,
1094                                 const struct bpos *search,
1095                                 const struct bkey_packed *packed_search)
1096 {
1097         struct ro_aux_tree *base = ro_aux_tree_base(b, t);
1098         struct bkey_float *f;
1099         struct bkey_packed *k;
1100         unsigned inorder, n = 1, l, r;
1101         int cmp;
1102
1103         do {
1104                 if (likely(n << 4 < t->size))
1105                         prefetch(&base->f[n << 4]);
1106
1107                 f = &base->f[n];
1108                 if (unlikely(f->exponent >= BFLOAT_FAILED))
1109                         goto slowpath;
1110
1111                 l = f->mantissa;
1112                 r = bkey_mantissa(packed_search, f);
1113
1114                 if (unlikely(l == r) && bkey_mantissa_bits_dropped(b, f))
1115                         goto slowpath;
1116
1117                 n = n * 2 + (l < r);
1118                 continue;
1119 slowpath:
1120                 k = tree_to_bkey(b, t, n);
1121                 cmp = bkey_cmp_p_or_unp(b, k, packed_search, search);
1122                 if (!cmp)
1123                         return k;
1124
1125                 n = n * 2 + (cmp < 0);
1126         } while (n < t->size);
1127
1128         inorder = __eytzinger1_to_inorder(n >> 1, t->size - 1, t->extra);
1129
1130         /*
1131          * n would have been the node we recursed to - the low bit tells us if
1132          * we recursed left or recursed right.
1133          */
1134         if (likely(!(n & 1))) {
1135                 --inorder;
1136                 if (unlikely(!inorder))
1137                         return btree_bkey_first(b, t);
1138
1139                 f = &base->f[eytzinger1_prev(n >> 1, t->size - 1)];
1140         }
1141
1142         return cacheline_to_bkey(b, t, inorder, f->key_offset);
1143 }
1144
1145 static __always_inline __flatten
1146 struct bkey_packed *__bch2_bset_search(struct btree *b,
1147                                 struct bset_tree *t,
1148                                 struct bpos *search,
1149                                 const struct bkey_packed *lossy_packed_search)
1150 {
1151
1152         /*
1153          * First, we search for a cacheline, then lastly we do a linear search
1154          * within that cacheline.
1155          *
1156          * To search for the cacheline, there's three different possibilities:
1157          *  * The set is too small to have a search tree, so we just do a linear
1158          *    search over the whole set.
1159          *  * The set is the one we're currently inserting into; keeping a full
1160          *    auxiliary search tree up to date would be too expensive, so we
1161          *    use a much simpler lookup table to do a binary search -
1162          *    bset_search_write_set().
1163          *  * Or we use the auxiliary search tree we constructed earlier -
1164          *    bset_search_tree()
1165          */
1166
1167         switch (bset_aux_tree_type(t)) {
1168         case BSET_NO_AUX_TREE:
1169                 return btree_bkey_first(b, t);
1170         case BSET_RW_AUX_TREE:
1171                 return bset_search_write_set(b, t, search);
1172         case BSET_RO_AUX_TREE:
1173                 return bset_search_tree(b, t, search, lossy_packed_search);
1174         default:
1175                 BUG();
1176         }
1177 }
1178
1179 static __always_inline __flatten
1180 struct bkey_packed *bch2_bset_search_linear(struct btree *b,
1181                                 struct bset_tree *t,
1182                                 struct bpos *search,
1183                                 struct bkey_packed *packed_search,
1184                                 const struct bkey_packed *lossy_packed_search,
1185                                 struct bkey_packed *m)
1186 {
1187         if (lossy_packed_search)
1188                 while (m != btree_bkey_last(b, t) &&
1189                        bkey_iter_cmp_p_or_unp(b, m,
1190                                         lossy_packed_search, search) < 0)
1191                         m = bkey_p_next(m);
1192
1193         if (!packed_search)
1194                 while (m != btree_bkey_last(b, t) &&
1195                        bkey_iter_pos_cmp(b, m, search) < 0)
1196                         m = bkey_p_next(m);
1197
1198         if (bch2_expensive_debug_checks) {
1199                 struct bkey_packed *prev = bch2_bkey_prev_all(b, t, m);
1200
1201                 BUG_ON(prev &&
1202                        bkey_iter_cmp_p_or_unp(b, prev,
1203                                         packed_search, search) >= 0);
1204         }
1205
1206         return m;
1207 }
1208
1209 /* Btree node iterator */
1210
1211 static inline void __bch2_btree_node_iter_push(struct btree_node_iter *iter,
1212                               struct btree *b,
1213                               const struct bkey_packed *k,
1214                               const struct bkey_packed *end)
1215 {
1216         if (k != end) {
1217                 struct btree_node_iter_set *pos;
1218
1219                 btree_node_iter_for_each(iter, pos)
1220                         ;
1221
1222                 BUG_ON(pos >= iter->data + ARRAY_SIZE(iter->data));
1223                 *pos = (struct btree_node_iter_set) {
1224                         __btree_node_key_to_offset(b, k),
1225                         __btree_node_key_to_offset(b, end)
1226                 };
1227         }
1228 }
1229
1230 void bch2_btree_node_iter_push(struct btree_node_iter *iter,
1231                                struct btree *b,
1232                                const struct bkey_packed *k,
1233                                const struct bkey_packed *end)
1234 {
1235         __bch2_btree_node_iter_push(iter, b, k, end);
1236         bch2_btree_node_iter_sort(iter, b);
1237 }
1238
1239 noinline __flatten __cold
1240 static void btree_node_iter_init_pack_failed(struct btree_node_iter *iter,
1241                               struct btree *b, struct bpos *search)
1242 {
1243         struct bkey_packed *k;
1244
1245         trace_bkey_pack_pos_fail(search);
1246
1247         bch2_btree_node_iter_init_from_start(iter, b);
1248
1249         while ((k = bch2_btree_node_iter_peek(iter, b)) &&
1250                bkey_iter_pos_cmp(b, k, search) < 0)
1251                 bch2_btree_node_iter_advance(iter, b);
1252 }
1253
1254 /**
1255  * bch2_btree_node_iter_init - initialize a btree node iterator, starting from a
1256  * given position
1257  *
1258  * @iter:       iterator to initialize
1259  * @b:          btree node to search
1260  * @search:     search key
1261  *
1262  * Main entry point to the lookup code for individual btree nodes:
1263  *
1264  * NOTE:
1265  *
1266  * When you don't filter out deleted keys, btree nodes _do_ contain duplicate
1267  * keys. This doesn't matter for most code, but it does matter for lookups.
1268  *
1269  * Some adjacent keys with a string of equal keys:
1270  *      i j k k k k l m
1271  *
1272  * If you search for k, the lookup code isn't guaranteed to return you any
1273  * specific k. The lookup code is conceptually doing a binary search and
1274  * iterating backwards is very expensive so if the pivot happens to land at the
1275  * last k that's what you'll get.
1276  *
1277  * This works out ok, but it's something to be aware of:
1278  *
1279  *  - For non extents, we guarantee that the live key comes last - see
1280  *    btree_node_iter_cmp(), keys_out_of_order(). So the duplicates you don't
1281  *    see will only be deleted keys you don't care about.
1282  *
1283  *  - For extents, deleted keys sort last (see the comment at the top of this
1284  *    file). But when you're searching for extents, you actually want the first
1285  *    key strictly greater than your search key - an extent that compares equal
1286  *    to the search key is going to have 0 sectors after the search key.
1287  *
1288  *    But this does mean that we can't just search for
1289  *    bpos_successor(start_of_range) to get the first extent that overlaps with
1290  *    the range we want - if we're unlucky and there's an extent that ends
1291  *    exactly where we searched, then there could be a deleted key at the same
1292  *    position and we'd get that when we search instead of the preceding extent
1293  *    we needed.
1294  *
1295  *    So we've got to search for start_of_range, then after the lookup iterate
1296  *    past any extents that compare equal to the position we searched for.
1297  */
1298 __flatten
1299 void bch2_btree_node_iter_init(struct btree_node_iter *iter,
1300                                struct btree *b, struct bpos *search)
1301 {
1302         struct bkey_packed p, *packed_search = NULL;
1303         struct btree_node_iter_set *pos = iter->data;
1304         struct bkey_packed *k[MAX_BSETS];
1305         unsigned i;
1306
1307         EBUG_ON(bpos_lt(*search, b->data->min_key));
1308         EBUG_ON(bpos_gt(*search, b->data->max_key));
1309         bset_aux_tree_verify(b);
1310
1311         memset(iter, 0, sizeof(*iter));
1312
1313         switch (bch2_bkey_pack_pos_lossy(&p, *search, b)) {
1314         case BKEY_PACK_POS_EXACT:
1315                 packed_search = &p;
1316                 break;
1317         case BKEY_PACK_POS_SMALLER:
1318                 packed_search = NULL;
1319                 break;
1320         case BKEY_PACK_POS_FAIL:
1321                 btree_node_iter_init_pack_failed(iter, b, search);
1322                 return;
1323         }
1324
1325         for (i = 0; i < b->nsets; i++) {
1326                 k[i] = __bch2_bset_search(b, b->set + i, search, &p);
1327                 prefetch_four_cachelines(k[i]);
1328         }
1329
1330         for (i = 0; i < b->nsets; i++) {
1331                 struct bset_tree *t = b->set + i;
1332                 struct bkey_packed *end = btree_bkey_last(b, t);
1333
1334                 k[i] = bch2_bset_search_linear(b, t, search,
1335                                                packed_search, &p, k[i]);
1336                 if (k[i] != end)
1337                         *pos++ = (struct btree_node_iter_set) {
1338                                 __btree_node_key_to_offset(b, k[i]),
1339                                 __btree_node_key_to_offset(b, end)
1340                         };
1341         }
1342
1343         bch2_btree_node_iter_sort(iter, b);
1344 }
1345
1346 void bch2_btree_node_iter_init_from_start(struct btree_node_iter *iter,
1347                                           struct btree *b)
1348 {
1349         memset(iter, 0, sizeof(*iter));
1350
1351         for_each_bset(b, t)
1352                 __bch2_btree_node_iter_push(iter, b,
1353                                            btree_bkey_first(b, t),
1354                                            btree_bkey_last(b, t));
1355         bch2_btree_node_iter_sort(iter, b);
1356 }
1357
1358 struct bkey_packed *bch2_btree_node_iter_bset_pos(struct btree_node_iter *iter,
1359                                                   struct btree *b,
1360                                                   struct bset_tree *t)
1361 {
1362         struct btree_node_iter_set *set;
1363
1364         btree_node_iter_for_each(iter, set)
1365                 if (set->end == t->end_offset)
1366                         return __btree_node_offset_to_key(b, set->k);
1367
1368         return btree_bkey_last(b, t);
1369 }
1370
1371 static inline bool btree_node_iter_sort_two(struct btree_node_iter *iter,
1372                                             struct btree *b,
1373                                             unsigned first)
1374 {
1375         bool ret;
1376
1377         if ((ret = (btree_node_iter_cmp(b,
1378                                         iter->data[first],
1379                                         iter->data[first + 1]) > 0)))
1380                 swap(iter->data[first], iter->data[first + 1]);
1381         return ret;
1382 }
1383
1384 void bch2_btree_node_iter_sort(struct btree_node_iter *iter,
1385                                struct btree *b)
1386 {
1387         /* unrolled bubble sort: */
1388
1389         if (!__btree_node_iter_set_end(iter, 2)) {
1390                 btree_node_iter_sort_two(iter, b, 0);
1391                 btree_node_iter_sort_two(iter, b, 1);
1392         }
1393
1394         if (!__btree_node_iter_set_end(iter, 1))
1395                 btree_node_iter_sort_two(iter, b, 0);
1396 }
1397
1398 void bch2_btree_node_iter_set_drop(struct btree_node_iter *iter,
1399                                    struct btree_node_iter_set *set)
1400 {
1401         struct btree_node_iter_set *last =
1402                 iter->data + ARRAY_SIZE(iter->data) - 1;
1403
1404         memmove(&set[0], &set[1], (void *) last - (void *) set);
1405         *last = (struct btree_node_iter_set) { 0, 0 };
1406 }
1407
1408 static inline void __bch2_btree_node_iter_advance(struct btree_node_iter *iter,
1409                                                   struct btree *b)
1410 {
1411         iter->data->k += __bch2_btree_node_iter_peek_all(iter, b)->u64s;
1412
1413         EBUG_ON(iter->data->k > iter->data->end);
1414
1415         if (unlikely(__btree_node_iter_set_end(iter, 0))) {
1416                 /* avoid an expensive memmove call: */
1417                 iter->data[0] = iter->data[1];
1418                 iter->data[1] = iter->data[2];
1419                 iter->data[2] = (struct btree_node_iter_set) { 0, 0 };
1420                 return;
1421         }
1422
1423         if (__btree_node_iter_set_end(iter, 1))
1424                 return;
1425
1426         if (!btree_node_iter_sort_two(iter, b, 0))
1427                 return;
1428
1429         if (__btree_node_iter_set_end(iter, 2))
1430                 return;
1431
1432         btree_node_iter_sort_two(iter, b, 1);
1433 }
1434
1435 void bch2_btree_node_iter_advance(struct btree_node_iter *iter,
1436                                   struct btree *b)
1437 {
1438         if (bch2_expensive_debug_checks) {
1439                 bch2_btree_node_iter_verify(iter, b);
1440                 bch2_btree_node_iter_next_check(iter, b);
1441         }
1442
1443         __bch2_btree_node_iter_advance(iter, b);
1444 }
1445
1446 /*
1447  * Expensive:
1448  */
1449 struct bkey_packed *bch2_btree_node_iter_prev_all(struct btree_node_iter *iter,
1450                                                   struct btree *b)
1451 {
1452         struct bkey_packed *k, *prev = NULL;
1453         struct btree_node_iter_set *set;
1454         unsigned end = 0;
1455
1456         if (bch2_expensive_debug_checks)
1457                 bch2_btree_node_iter_verify(iter, b);
1458
1459         for_each_bset(b, t) {
1460                 k = bch2_bkey_prev_all(b, t,
1461                         bch2_btree_node_iter_bset_pos(iter, b, t));
1462                 if (k &&
1463                     (!prev || bkey_iter_cmp(b, k, prev) > 0)) {
1464                         prev = k;
1465                         end = t->end_offset;
1466                 }
1467         }
1468
1469         if (!prev)
1470                 return NULL;
1471
1472         /*
1473          * We're manually memmoving instead of just calling sort() to ensure the
1474          * prev we picked ends up in slot 0 - sort won't necessarily put it
1475          * there because of duplicate deleted keys:
1476          */
1477         btree_node_iter_for_each(iter, set)
1478                 if (set->end == end)
1479                         goto found;
1480
1481         BUG_ON(set != &iter->data[__btree_node_iter_used(iter)]);
1482 found:
1483         BUG_ON(set >= iter->data + ARRAY_SIZE(iter->data));
1484
1485         memmove(&iter->data[1],
1486                 &iter->data[0],
1487                 (void *) set - (void *) &iter->data[0]);
1488
1489         iter->data[0].k = __btree_node_key_to_offset(b, prev);
1490         iter->data[0].end = end;
1491
1492         if (bch2_expensive_debug_checks)
1493                 bch2_btree_node_iter_verify(iter, b);
1494         return prev;
1495 }
1496
1497 struct bkey_packed *bch2_btree_node_iter_prev(struct btree_node_iter *iter,
1498                                               struct btree *b)
1499 {
1500         struct bkey_packed *prev;
1501
1502         do {
1503                 prev = bch2_btree_node_iter_prev_all(iter, b);
1504         } while (prev && bkey_deleted(prev));
1505
1506         return prev;
1507 }
1508
1509 struct bkey_s_c bch2_btree_node_iter_peek_unpack(struct btree_node_iter *iter,
1510                                                  struct btree *b,
1511                                                  struct bkey *u)
1512 {
1513         struct bkey_packed *k = bch2_btree_node_iter_peek(iter, b);
1514
1515         return k ? bkey_disassemble(b, k, u) : bkey_s_c_null;
1516 }
1517
1518 /* Mergesort */
1519
1520 void bch2_btree_keys_stats(const struct btree *b, struct bset_stats *stats)
1521 {
1522         for_each_bset_c(b, t) {
1523                 enum bset_aux_tree_type type = bset_aux_tree_type(t);
1524                 size_t j;
1525
1526                 stats->sets[type].nr++;
1527                 stats->sets[type].bytes += le16_to_cpu(bset(b, t)->u64s) *
1528                         sizeof(u64);
1529
1530                 if (bset_has_ro_aux_tree(t)) {
1531                         stats->floats += t->size - 1;
1532
1533                         for (j = 1; j < t->size; j++)
1534                                 stats->failed +=
1535                                         bkey_float(b, t, j)->exponent ==
1536                                         BFLOAT_FAILED;
1537                 }
1538         }
1539 }
1540
1541 void bch2_bfloat_to_text(struct printbuf *out, struct btree *b,
1542                          struct bkey_packed *k)
1543 {
1544         struct bset_tree *t = bch2_bkey_to_bset(b, k);
1545         struct bkey uk;
1546         unsigned j, inorder;
1547
1548         if (!bset_has_ro_aux_tree(t))
1549                 return;
1550
1551         inorder = bkey_to_cacheline(b, t, k);
1552         if (!inorder || inorder >= t->size)
1553                 return;
1554
1555         j = __inorder_to_eytzinger1(inorder, t->size - 1, t->extra);
1556         if (k != tree_to_bkey(b, t, j))
1557                 return;
1558
1559         switch (bkey_float(b, t, j)->exponent) {
1560         case BFLOAT_FAILED:
1561                 uk = bkey_unpack_key(b, k);
1562                 prt_printf(out,
1563                        "    failed unpacked at depth %u\n"
1564                        "\t",
1565                        ilog2(j));
1566                 bch2_bpos_to_text(out, uk.p);
1567                 prt_printf(out, "\n");
1568                 break;
1569         }
1570 }