git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/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 "util.h"
  14
  15 #include <asm/unaligned.h>
  16 #include <linux/console.h>
  17 #include <linux/random.h>
  18 #include <linux/prefetch.h>
  19
  20 /* hack.. */
  21 #include "alloc_types.h"
  22 #include <trace/events/bcachefs.h>
  23
  24 static inline void __bch2_btree_node_iter_advance(struct btree_node_iter *,
  25                                                   struct btree *);
  26
  27 static inline unsigned __btree_node_iter_used(struct btree_node_iter *iter)
  28 {
  29         unsigned n = ARRAY_SIZE(iter->data);
  30
  31         while (n && __btree_node_iter_set_end(iter, n - 1))
  32                 --n;
  33
  34         return n;
  35 }
  36
  37 struct bset_tree *bch2_bkey_to_bset(struct btree *b, struct bkey_packed *k)
  38 {
  39         return bch2_bkey_to_bset_inlined(b, k);
  40 }
  41
  42 /*
  43  * There are never duplicate live keys in the btree - but including keys that
  44  * have been flagged as deleted (and will be cleaned up later) we _will_ see
  45  * duplicates.
  46  *
  47  * Thus the sort order is: usual key comparison first, but for keys that compare
  48  * equal the deleted key(s) come first, and the (at most one) live version comes
  49  * last.
  50  *
  51  * The main reason for this is insertion: to handle overwrites, we first iterate
  52  * over keys that compare equal to our insert key, and then insert immediately
  53  * prior to the first key greater than the key we're inserting - our insert
  54  * position will be after all keys that compare equal to our insert key, which
  55  * by the time we actually do the insert will all be deleted.
  56  */
  57
  58 void bch2_dump_bset(struct bch_fs *c, struct btree *b,
  59                     struct bset *i, unsigned set)
  60 {
  61         struct bkey_packed *_k, *_n;
  62         struct bkey uk, n;
  63         struct bkey_s_c k;
  64         struct printbuf buf = PRINTBUF;
  65
  66         if (!i->u64s)
  67                 return;
  68
  69         for (_k = i->start;
  70              _k < vstruct_last(i);
  71              _k = _n) {
  72                 _n = bkey_p_next(_k);
  73
  74                 k = bkey_disassemble(b, _k, &uk);
  75
  76                 printbuf_reset(&buf);
  77                 if (c)
  78                         bch2_bkey_val_to_text(&buf, c, k);
  79                 else
  80                         bch2_bkey_to_text(&buf, k.k);
  81                 printk(KERN_ERR "block %u key %5zu: %s\n", set,
  82                        _k->_data - i->_data, buf.buf);
  83
  84                 if (_n == vstruct_last(i))
  85                         continue;
  86
  87                 n = bkey_unpack_key(b, _n);
  88
  89                 if (bpos_lt(n.p, k.k->p)) {
  90                         printk(KERN_ERR "Key skipped backwards\n");
  91                         continue;
  92                 }
  93
  94                 if (!bkey_deleted(k.k) && bpos_eq(n.p, k.k->p))
  95                         printk(KERN_ERR "Duplicate keys\n");
  96         }
  97
  98         printbuf_exit(&buf);
  99 }
 100
 101 void bch2_dump_btree_node(struct bch_fs *c, struct btree *b)
 102 {
 103         struct bset_tree *t;
 104
 105         console_lock();
 106         for_each_bset(b, t)
 107                 bch2_dump_bset(c, b, bset(b, t), t - b->set);
 108         console_unlock();
 109 }
 110
 111 void bch2_dump_btree_node_iter(struct btree *b,
 112                               struct btree_node_iter *iter)
 113 {
 114         struct btree_node_iter_set *set;
 115         struct printbuf buf = PRINTBUF;
 116
 117         printk(KERN_ERR "btree node iter with %u/%u sets:\n",
 118                __btree_node_iter_used(iter), b->nsets);
 119
 120         btree_node_iter_for_each(iter, set) {
 121                 struct bkey_packed *k = __btree_node_offset_to_key(b, set->k);
 122                 struct bset_tree *t = bch2_bkey_to_bset(b, k);
 123                 struct bkey uk = bkey_unpack_key(b, k);
 124
 125                 printbuf_reset(&buf);
 126                 bch2_bkey_to_text(&buf, &uk);
 127                 printk(KERN_ERR "set %zu key %u: %s\n",
 128                        t - b->set, set->k, buf.buf);
 129         }
 130
 131         printbuf_exit(&buf);
 132 }
 133
 134 #ifdef CONFIG_BCACHEFS_DEBUG
 135
 136 void __bch2_verify_btree_nr_keys(struct btree *b)
 137 {
 138         struct bset_tree *t;
 139         struct bkey_packed *k;
 140         struct btree_nr_keys nr = { 0 };
 141
 142         for_each_bset(b, t)
 143                 bset_tree_for_each_key(b, t, k)
 144                         if (!bkey_deleted(k))
 145                                 btree_keys_account_key_add(&nr, t - b->set, k);
 146
 147         BUG_ON(memcmp(&nr, &b->nr, sizeof(nr)));
 148 }
 149
 150 static void bch2_btree_node_iter_next_check(struct btree_node_iter *_iter,
 151                                             struct btree *b)
 152 {
 153         struct btree_node_iter iter = *_iter;
 154         const struct bkey_packed *k, *n;
 155
 156         k = bch2_btree_node_iter_peek_all(&iter, b);
 157         __bch2_btree_node_iter_advance(&iter, b);
 158         n = bch2_btree_node_iter_peek_all(&iter, b);
 159
 160         bkey_unpack_key(b, k);
 161
 162         if (n &&
 163             bkey_iter_cmp(b, k, n) > 0) {
 164                 struct btree_node_iter_set *set;
 165                 struct bkey ku = bkey_unpack_key(b, k);
 166                 struct bkey nu = bkey_unpack_key(b, n);
 167                 struct printbuf buf1 = PRINTBUF;
 168                 struct printbuf buf2 = PRINTBUF;
 169
 170                 bch2_dump_btree_node(NULL, b);
 171                 bch2_bkey_to_text(&buf1, &ku);
 172                 bch2_bkey_to_text(&buf2, &nu);
 173                 printk(KERN_ERR "out of order/overlapping:\n%s\n%s\n",
 174                        buf1.buf, buf2.buf);
 175                 printk(KERN_ERR "iter was:");
 176
 177                 btree_node_iter_for_each(_iter, set) {
 178                         struct bkey_packed *k = __btree_node_offset_to_key(b, set->k);
 179                         struct bset_tree *t = bch2_bkey_to_bset(b, k);
 180                         printk(" [%zi %zi]", t - b->set,
 181                                k->_data - bset(b, t)->_data);
 182                 }
 183                 panic("\n");
 184         }
 185 }
 186
 187 void bch2_btree_node_iter_verify(struct btree_node_iter *iter,
 188                                  struct btree *b)
 189 {
 190         struct btree_node_iter_set *set, *s2;
 191         struct bkey_packed *k, *p;
 192         struct bset_tree *t;
 193
 194         if (bch2_btree_node_iter_end(iter))
 195                 return;
 196
 197         /* Verify no duplicates: */
 198         btree_node_iter_for_each(iter, set) {
 199                 BUG_ON(set->k > set->end);
 200                 btree_node_iter_for_each(iter, s2)
 201                         BUG_ON(set != s2 && set->end == s2->end);
 202         }
 203
 204         /* Verify that set->end is correct: */
 205         btree_node_iter_for_each(iter, set) {
 206                 for_each_bset(b, t)
 207                         if (set->end == t->end_offset)
 208                                 goto found;
 209                 BUG();
 210 found:
 211                 BUG_ON(set->k < btree_bkey_first_offset(t) ||
 212                        set->k >= t->end_offset);
 213         }
 214
 215         /* Verify iterator is sorted: */
 216         btree_node_iter_for_each(iter, set)
 217                 BUG_ON(set != iter->data &&
 218                        btree_node_iter_cmp(b, set[-1], set[0]) > 0);
 219
 220         k = bch2_btree_node_iter_peek_all(iter, b);
 221
 222         for_each_bset(b, t) {
 223                 if (iter->data[0].end == t->end_offset)
 224                         continue;
 225
 226                 p = bch2_bkey_prev_all(b, t,
 227                         bch2_btree_node_iter_bset_pos(iter, b, t));
 228
 229                 BUG_ON(p && bkey_iter_cmp(b, k, p) < 0);
 230         }
 231 }
 232
 233 void bch2_verify_insert_pos(struct btree *b, struct bkey_packed *where,
 234                             struct bkey_packed *insert, unsigned clobber_u64s)
 235 {
 236         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 237         struct bkey_packed *prev = bch2_bkey_prev_all(b, t, where);
 238         struct bkey_packed *next = (void *) (where->_data + clobber_u64s);
 239         struct printbuf buf1 = PRINTBUF;
 240         struct printbuf buf2 = PRINTBUF;
 241 #if 0
 242         BUG_ON(prev &&
 243                bkey_iter_cmp(b, prev, insert) > 0);
 244 #else
 245         if (prev &&
 246             bkey_iter_cmp(b, prev, insert) > 0) {
 247                 struct bkey k1 = bkey_unpack_key(b, prev);
 248                 struct bkey k2 = bkey_unpack_key(b, insert);
 249
 250                 bch2_dump_btree_node(NULL, b);
 251                 bch2_bkey_to_text(&buf1, &k1);
 252                 bch2_bkey_to_text(&buf2, &k2);
 253
 254                 panic("prev > insert:\n"
 255                       "prev    key %s\n"
 256                       "insert  key %s\n",
 257                       buf1.buf, buf2.buf);
 258         }
 259 #endif
 260 #if 0
 261         BUG_ON(next != btree_bkey_last(b, t) &&
 262                bkey_iter_cmp(b, insert, next) > 0);
 263 #else
 264         if (next != btree_bkey_last(b, t) &&
 265             bkey_iter_cmp(b, insert, next) > 0) {
 266                 struct bkey k1 = bkey_unpack_key(b, insert);
 267                 struct bkey k2 = bkey_unpack_key(b, next);
 268
 269                 bch2_dump_btree_node(NULL, b);
 270                 bch2_bkey_to_text(&buf1, &k1);
 271                 bch2_bkey_to_text(&buf2, &k2);
 272
 273                 panic("insert > next:\n"
 274                       "insert  key %s\n"
 275                       "next    key %s\n",
 276                       buf1.buf, buf2.buf);
 277         }
 278 #endif
 279 }
 280
 281 #else
 282
 283 static inline void bch2_btree_node_iter_next_check(struct btree_node_iter *iter,
 284                                                    struct btree *b) {}
 285
 286 #endif
 287
 288 /* Auxiliary search trees */
 289
 290 #define BFLOAT_FAILED_UNPACKED  U8_MAX
 291 #define BFLOAT_FAILED           U8_MAX
 292
 293 struct bkey_float {
 294         u8              exponent;
 295         u8              key_offset;
 296         u16             mantissa;
 297 };
 298 #define BKEY_MANTISSA_BITS      16
 299
 300 static unsigned bkey_float_byte_offset(unsigned idx)
 301 {
 302         return idx * sizeof(struct bkey_float);
 303 }
 304
 305 struct ro_aux_tree {
 306         struct bkey_float       f[0];
 307 };
 308
 309 struct rw_aux_tree {
 310         u16             offset;
 311         struct bpos     k;
 312 };
 313
 314 static unsigned bset_aux_tree_buf_end(const struct bset_tree *t)
 315 {
 316         BUG_ON(t->aux_data_offset == U16_MAX);
 317
 318         switch (bset_aux_tree_type(t)) {
 319         case BSET_NO_AUX_TREE:
 320                 return t->aux_data_offset;
 321         case BSET_RO_AUX_TREE:
 322                 return t->aux_data_offset +
 323                         DIV_ROUND_UP(t->size * sizeof(struct bkey_float) +
 324                                      t->size * sizeof(u8), 8);
 325         case BSET_RW_AUX_TREE:
 326                 return t->aux_data_offset +
 327                         DIV_ROUND_UP(sizeof(struct rw_aux_tree) * t->size, 8);
 328         default:
 329                 BUG();
 330         }
 331 }
 332
 333 static unsigned bset_aux_tree_buf_start(const struct btree *b,
 334                                         const struct bset_tree *t)
 335 {
 336         return t == b->set
 337                 ? DIV_ROUND_UP(b->unpack_fn_len, 8)
 338                 : bset_aux_tree_buf_end(t - 1);
 339 }
 340
 341 static void *__aux_tree_base(const struct btree *b,
 342                              const struct bset_tree *t)
 343 {
 344         return b->aux_data + t->aux_data_offset * 8;
 345 }
 346
 347 static struct ro_aux_tree *ro_aux_tree_base(const struct btree *b,
 348                                             const struct bset_tree *t)
 349 {
 350         EBUG_ON(bset_aux_tree_type(t) != BSET_RO_AUX_TREE);
 351
 352         return __aux_tree_base(b, t);
 353 }
 354
 355 static u8 *ro_aux_tree_prev(const struct btree *b,
 356                             const struct bset_tree *t)
 357 {
 358         EBUG_ON(bset_aux_tree_type(t) != BSET_RO_AUX_TREE);
 359
 360         return __aux_tree_base(b, t) + bkey_float_byte_offset(t->size);
 361 }
 362
 363 static struct bkey_float *bkey_float(const struct btree *b,
 364                                      const struct bset_tree *t,
 365                                      unsigned idx)
 366 {
 367         return ro_aux_tree_base(b, t)->f + idx;
 368 }
 369
 370 static void bset_aux_tree_verify(const struct btree *b)
 371 {
 372 #ifdef CONFIG_BCACHEFS_DEBUG
 373         const struct bset_tree *t;
 374
 375         for_each_bset(b, t) {
 376                 if (t->aux_data_offset == U16_MAX)
 377                         continue;
 378
 379                 BUG_ON(t != b->set &&
 380                        t[-1].aux_data_offset == U16_MAX);
 381
 382                 BUG_ON(t->aux_data_offset < bset_aux_tree_buf_start(b, t));
 383                 BUG_ON(t->aux_data_offset > btree_aux_data_u64s(b));
 384                 BUG_ON(bset_aux_tree_buf_end(t) > btree_aux_data_u64s(b));
 385         }
 386 #endif
 387 }
 388
 389 void bch2_btree_keys_init(struct btree *b)
 390 {
 391         unsigned i;
 392
 393         b->nsets                = 0;
 394         memset(&b->nr, 0, sizeof(b->nr));
 395
 396         for (i = 0; i < MAX_BSETS; i++)
 397                 b->set[i].data_offset = U16_MAX;
 398
 399         bch2_bset_set_no_aux_tree(b, b->set);
 400 }
 401
 402 /* Binary tree stuff for auxiliary search trees */
 403
 404 /*
 405  * Cacheline/offset <-> bkey pointer arithmetic:
 406  *
 407  * t->tree is a binary search tree in an array; each node corresponds to a key
 408  * in one cacheline in t->set (BSET_CACHELINE bytes).
 409  *
 410  * This means we don't have to store the full index of the key that a node in
 411  * the binary tree points to; eytzinger1_to_inorder() gives us the cacheline, and
 412  * then bkey_float->m gives us the offset within that cacheline, in units of 8
 413  * bytes.
 414  *
 415  * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
 416  * make this work.
 417  *
 418  * To construct the bfloat for an arbitrary key we need to know what the key
 419  * immediately preceding it is: we have to check if the two keys differ in the
 420  * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
 421  * of the previous key so we can walk backwards to it from t->tree[j]'s key.
 422  */
 423
 424 static inline void *bset_cacheline(const struct btree *b,
 425                                    const struct bset_tree *t,
 426                                    unsigned cacheline)
 427 {
 428         return (void *) round_down((unsigned long) btree_bkey_first(b, t),
 429                                    L1_CACHE_BYTES) +
 430                 cacheline * BSET_CACHELINE;
 431 }
 432
 433 static struct bkey_packed *cacheline_to_bkey(const struct btree *b,
 434                                              const struct bset_tree *t,
 435                                              unsigned cacheline,
 436                                              unsigned offset)
 437 {
 438         return bset_cacheline(b, t, cacheline) + offset * 8;
 439 }
 440
 441 static unsigned bkey_to_cacheline(const struct btree *b,
 442                                   const struct bset_tree *t,
 443                                   const struct bkey_packed *k)
 444 {
 445         return ((void *) k - bset_cacheline(b, t, 0)) / BSET_CACHELINE;
 446 }
 447
 448 static ssize_t __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         return (u64 *) k - (u64 *) bset_cacheline(b, t, cacheline);
 454 }
 455
 456 static unsigned bkey_to_cacheline_offset(const struct btree *b,
 457                                          const struct bset_tree *t,
 458                                          unsigned cacheline,
 459                                          const struct bkey_packed *k)
 460 {
 461         size_t m = __bkey_to_cacheline_offset(b, t, cacheline, k);
 462
 463         EBUG_ON(m > U8_MAX);
 464         return m;
 465 }
 466
 467 static inline struct bkey_packed *tree_to_bkey(const struct btree *b,
 468                                                const struct bset_tree *t,
 469                                                unsigned j)
 470 {
 471         return cacheline_to_bkey(b, t,
 472                         __eytzinger1_to_inorder(j, t->size - 1, t->extra),
 473                         bkey_float(b, t, j)->key_offset);
 474 }
 475
 476 static struct bkey_packed *tree_to_prev_bkey(const struct btree *b,
 477                                              const struct bset_tree *t,
 478                                              unsigned j)
 479 {
 480         unsigned prev_u64s = ro_aux_tree_prev(b, t)[j];
 481
 482         return (void *) (tree_to_bkey(b, t, j)->_data - prev_u64s);
 483 }
 484
 485 static struct rw_aux_tree *rw_aux_tree(const struct btree *b,
 486                                        const struct bset_tree *t)
 487 {
 488         EBUG_ON(bset_aux_tree_type(t) != BSET_RW_AUX_TREE);
 489
 490         return __aux_tree_base(b, t);
 491 }
 492
 493 /*
 494  * For the write set - the one we're currently inserting keys into - we don't
 495  * maintain a full search tree, we just keep a simple lookup table in t->prev.
 496  */
 497 static struct bkey_packed *rw_aux_to_bkey(const struct btree *b,
 498                                           struct bset_tree *t,
 499                                           unsigned j)
 500 {
 501         return __btree_node_offset_to_key(b, rw_aux_tree(b, t)[j].offset);
 502 }
 503
 504 static void rw_aux_tree_set(const struct btree *b, struct bset_tree *t,
 505                             unsigned j, struct bkey_packed *k)
 506 {
 507         EBUG_ON(k >= btree_bkey_last(b, t));
 508
 509         rw_aux_tree(b, t)[j] = (struct rw_aux_tree) {
 510                 .offset = __btree_node_key_to_offset(b, k),
 511                 .k      = bkey_unpack_pos(b, k),
 512         };
 513 }
 514
 515 static void bch2_bset_verify_rw_aux_tree(struct btree *b,
 516                                         struct bset_tree *t)
 517 {
 518         struct bkey_packed *k = btree_bkey_first(b, t);
 519         unsigned j = 0;
 520
 521         if (!bch2_expensive_debug_checks)
 522                 return;
 523
 524         BUG_ON(bset_has_ro_aux_tree(t));
 525
 526         if (!bset_has_rw_aux_tree(t))
 527                 return;
 528
 529         BUG_ON(t->size < 1);
 530         BUG_ON(rw_aux_to_bkey(b, t, j) != k);
 531
 532         goto start;
 533         while (1) {
 534                 if (rw_aux_to_bkey(b, t, j) == k) {
 535                         BUG_ON(!bpos_eq(rw_aux_tree(b, t)[j].k,
 536                                         bkey_unpack_pos(b, k)));
 537 start:
 538                         if (++j == t->size)
 539                                 break;
 540
 541                         BUG_ON(rw_aux_tree(b, t)[j].offset <=
 542                                rw_aux_tree(b, t)[j - 1].offset);
 543                 }
 544
 545                 k = bkey_p_next(k);
 546                 BUG_ON(k >= btree_bkey_last(b, t));
 547         }
 548 }
 549
 550 /* returns idx of first entry >= offset: */
 551 static unsigned rw_aux_tree_bsearch(struct btree *b,
 552                                     struct bset_tree *t,
 553                                     unsigned offset)
 554 {
 555         unsigned bset_offs = offset - btree_bkey_first_offset(t);
 556         unsigned bset_u64s = t->end_offset - btree_bkey_first_offset(t);
 557         unsigned idx = bset_u64s ? bset_offs * t->size / bset_u64s : 0;
 558
 559         EBUG_ON(bset_aux_tree_type(t) != BSET_RW_AUX_TREE);
 560         EBUG_ON(!t->size);
 561         EBUG_ON(idx > t->size);
 562
 563         while (idx < t->size &&
 564                rw_aux_tree(b, t)[idx].offset < offset)
 565                 idx++;
 566
 567         while (idx &&
 568                rw_aux_tree(b, t)[idx - 1].offset >= offset)
 569                 idx--;
 570
 571         EBUG_ON(idx < t->size &&
 572                 rw_aux_tree(b, t)[idx].offset < offset);
 573         EBUG_ON(idx && rw_aux_tree(b, t)[idx - 1].offset >= offset);
 574         EBUG_ON(idx + 1 < t->size &&
 575                 rw_aux_tree(b, t)[idx].offset ==
 576                 rw_aux_tree(b, t)[idx + 1].offset);
 577
 578         return idx;
 579 }
 580
 581 static inline unsigned bkey_mantissa(const struct bkey_packed *k,
 582                                      const struct bkey_float *f,
 583                                      unsigned idx)
 584 {
 585         u64 v;
 586
 587         EBUG_ON(!bkey_packed(k));
 588
 589         v = get_unaligned((u64 *) (((u8 *) k->_data) + (f->exponent >> 3)));
 590
 591         /*
 592          * In little endian, we're shifting off low bits (and then the bits we
 593          * want are at the low end), in big endian we're shifting off high bits
 594          * (and then the bits we want are at the high end, so we shift them
 595          * back down):
 596          */
 597 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 598         v >>= f->exponent & 7;
 599 #else
 600         v >>= 64 - (f->exponent & 7) - BKEY_MANTISSA_BITS;
 601 #endif
 602         return (u16) v;
 603 }
 604
 605 __always_inline
 606 static inline void make_bfloat(struct btree *b, struct bset_tree *t,
 607                                unsigned j,
 608                                struct bkey_packed *min_key,
 609                                struct bkey_packed *max_key)
 610 {
 611         struct bkey_float *f = bkey_float(b, t, j);
 612         struct bkey_packed *m = tree_to_bkey(b, t, j);
 613         struct bkey_packed *l = is_power_of_2(j)
 614                 ? min_key
 615                 : tree_to_prev_bkey(b, t, j >> ffs(j));
 616         struct bkey_packed *r = is_power_of_2(j + 1)
 617                 ? max_key
 618                 : tree_to_bkey(b, t, j >> (ffz(j) + 1));
 619         unsigned mantissa;
 620         int shift, exponent, high_bit;
 621
 622         /*
 623          * for failed bfloats, the lookup code falls back to comparing against
 624          * the original key.
 625          */
 626
 627         if (!bkey_packed(l) || !bkey_packed(r) || !bkey_packed(m) ||
 628             !b->nr_key_bits) {
 629                 f->exponent = BFLOAT_FAILED_UNPACKED;
 630                 return;
 631         }
 632
 633         /*
 634          * The greatest differing bit of l and r is the first bit we must
 635          * include in the bfloat mantissa we're creating in order to do
 636          * comparisons - that bit always becomes the high bit of
 637          * bfloat->mantissa, and thus the exponent we're calculating here is
 638          * the position of what will become the low bit in bfloat->mantissa:
 639          *
 640          * Note that this may be negative - we may be running off the low end
 641          * of the key: we handle this later:
 642          */
 643         high_bit = max(bch2_bkey_greatest_differing_bit(b, l, r),
 644                        min_t(unsigned, BKEY_MANTISSA_BITS, b->nr_key_bits) - 1);
 645         exponent = high_bit - (BKEY_MANTISSA_BITS - 1);
 646
 647         /*
 648          * Then we calculate the actual shift value, from the start of the key
 649          * (k->_data), to get the key bits starting at exponent:
 650          */
 651 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 652         shift = (int) (b->format.key_u64s * 64 - b->nr_key_bits) + exponent;
 653
 654         EBUG_ON(shift + BKEY_MANTISSA_BITS > b->format.key_u64s * 64);
 655 #else
 656         shift = high_bit_offset +
 657                 b->nr_key_bits -
 658                 exponent -
 659                 BKEY_MANTISSA_BITS;
 660
 661         EBUG_ON(shift < KEY_PACKED_BITS_START);
 662 #endif
 663         EBUG_ON(shift < 0 || shift >= BFLOAT_FAILED);
 664
 665         f->exponent = shift;
 666         mantissa = bkey_mantissa(m, f, j);
 667
 668         /*
 669          * If we've got garbage bits, set them to all 1s - it's legal for the
 670          * bfloat to compare larger than the original key, but not smaller:
 671          */
 672         if (exponent < 0)
 673                 mantissa |= ~(~0U << -exponent);
 674
 675         f->mantissa = mantissa;
 676 }
 677
 678 /* bytes remaining - only valid for last bset: */
 679 static unsigned __bset_tree_capacity(const struct btree *b, const struct bset_tree *t)
 680 {
 681         bset_aux_tree_verify(b);
 682
 683         return btree_aux_data_bytes(b) - t->aux_data_offset * sizeof(u64);
 684 }
 685
 686 static unsigned bset_ro_tree_capacity(const struct btree *b, const struct bset_tree *t)
 687 {
 688         return __bset_tree_capacity(b, t) /
 689                 (sizeof(struct bkey_float) + sizeof(u8));
 690 }
 691
 692 static unsigned bset_rw_tree_capacity(const struct btree *b, const struct bset_tree *t)
 693 {
 694         return __bset_tree_capacity(b, t) / sizeof(struct rw_aux_tree);
 695 }
 696
 697 static noinline void __build_rw_aux_tree(struct btree *b, struct bset_tree *t)
 698 {
 699         struct bkey_packed *k;
 700
 701         t->size = 1;
 702         t->extra = BSET_RW_AUX_TREE_VAL;
 703         rw_aux_tree(b, t)[0].offset =
 704                 __btree_node_key_to_offset(b, btree_bkey_first(b, t));
 705
 706         bset_tree_for_each_key(b, t, k) {
 707                 if (t->size == bset_rw_tree_capacity(b, t))
 708                         break;
 709
 710                 if ((void *) k - (void *) rw_aux_to_bkey(b, t, t->size - 1) >
 711                     L1_CACHE_BYTES)
 712                         rw_aux_tree_set(b, t, t->size++, k);
 713         }
 714 }
 715
 716 static noinline void __build_ro_aux_tree(struct btree *b, struct bset_tree *t)
 717 {
 718         struct bkey_packed *prev = NULL, *k = btree_bkey_first(b, t);
 719         struct bkey_i min_key, max_key;
 720         unsigned j, cacheline = 1;
 721
 722         t->size = min(bkey_to_cacheline(b, t, btree_bkey_last(b, t)),
 723                       bset_ro_tree_capacity(b, t));
 724 retry:
 725         if (t->size < 2) {
 726                 t->size = 0;
 727                 t->extra = BSET_NO_AUX_TREE_VAL;
 728                 return;
 729         }
 730
 731         t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
 732
 733         /* First we figure out where the first key in each cacheline is */
 734         eytzinger1_for_each(j, t->size - 1) {
 735                 while (bkey_to_cacheline(b, t, k) < cacheline)
 736                         prev = k, k = bkey_p_next(k);
 737
 738                 if (k >= btree_bkey_last(b, t)) {
 739                         /* XXX: this path sucks */
 740                         t->size--;
 741                         goto retry;
 742                 }
 743
 744                 ro_aux_tree_prev(b, t)[j] = prev->u64s;
 745                 bkey_float(b, t, j)->key_offset =
 746                         bkey_to_cacheline_offset(b, t, cacheline++, k);
 747
 748                 EBUG_ON(tree_to_prev_bkey(b, t, j) != prev);
 749                 EBUG_ON(tree_to_bkey(b, t, j) != k);
 750         }
 751
 752         while (k != btree_bkey_last(b, t))
 753                 prev = k, k = bkey_p_next(k);
 754
 755         if (!bkey_pack_pos(bkey_to_packed(&min_key), b->data->min_key, b)) {
 756                 bkey_init(&min_key.k);
 757                 min_key.k.p = b->data->min_key;
 758         }
 759
 760         if (!bkey_pack_pos(bkey_to_packed(&max_key), b->data->max_key, b)) {
 761                 bkey_init(&max_key.k);
 762                 max_key.k.p = b->data->max_key;
 763         }
 764
 765         /* Then we build the tree */
 766         eytzinger1_for_each(j, t->size - 1)
 767                 make_bfloat(b, t, j,
 768                             bkey_to_packed(&min_key),
 769                             bkey_to_packed(&max_key));
 770 }
 771
 772 static void bset_alloc_tree(struct btree *b, struct bset_tree *t)
 773 {
 774         struct bset_tree *i;
 775
 776         for (i = b->set; i != t; i++)
 777                 BUG_ON(bset_has_rw_aux_tree(i));
 778
 779         bch2_bset_set_no_aux_tree(b, t);
 780
 781         /* round up to next cacheline: */
 782         t->aux_data_offset = round_up(bset_aux_tree_buf_start(b, t),
 783                                       SMP_CACHE_BYTES / sizeof(u64));
 784
 785         bset_aux_tree_verify(b);
 786 }
 787
 788 void bch2_bset_build_aux_tree(struct btree *b, struct bset_tree *t,
 789                              bool writeable)
 790 {
 791         if (writeable
 792             ? bset_has_rw_aux_tree(t)
 793             : bset_has_ro_aux_tree(t))
 794                 return;
 795
 796         bset_alloc_tree(b, t);
 797
 798         if (!__bset_tree_capacity(b, t))
 799                 return;
 800
 801         if (writeable)
 802                 __build_rw_aux_tree(b, t);
 803         else
 804                 __build_ro_aux_tree(b, t);
 805
 806         bset_aux_tree_verify(b);
 807 }
 808
 809 void bch2_bset_init_first(struct btree *b, struct bset *i)
 810 {
 811         struct bset_tree *t;
 812
 813         BUG_ON(b->nsets);
 814
 815         memset(i, 0, sizeof(*i));
 816         get_random_bytes(&i->seq, sizeof(i->seq));
 817         SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 818
 819         t = &b->set[b->nsets++];
 820         set_btree_bset(b, t, i);
 821 }
 822
 823 void bch2_bset_init_next(struct bch_fs *c, struct btree *b,
 824                          struct btree_node_entry *bne)
 825 {
 826         struct bset *i = &bne->keys;
 827         struct bset_tree *t;
 828
 829         BUG_ON(bset_byte_offset(b, bne) >= btree_bytes(c));
 830         BUG_ON((void *) bne < (void *) btree_bkey_last(b, bset_tree_last(b)));
 831         BUG_ON(b->nsets >= MAX_BSETS);
 832
 833         memset(i, 0, sizeof(*i));
 834         i->seq = btree_bset_first(b)->seq;
 835         SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 836
 837         t = &b->set[b->nsets++];
 838         set_btree_bset(b, t, i);
 839 }
 840
 841 /*
 842  * find _some_ key in the same bset as @k that precedes @k - not necessarily the
 843  * immediate predecessor:
 844  */
 845 static struct bkey_packed *__bkey_prev(struct btree *b, struct bset_tree *t,
 846                                        struct bkey_packed *k)
 847 {
 848         struct bkey_packed *p;
 849         unsigned offset;
 850         int j;
 851
 852         EBUG_ON(k < btree_bkey_first(b, t) ||
 853                 k > btree_bkey_last(b, t));
 854
 855         if (k == btree_bkey_first(b, t))
 856                 return NULL;
 857
 858         switch (bset_aux_tree_type(t)) {
 859         case BSET_NO_AUX_TREE:
 860                 p = btree_bkey_first(b, t);
 861                 break;
 862         case BSET_RO_AUX_TREE:
 863                 j = min_t(unsigned, t->size - 1, bkey_to_cacheline(b, t, k));
 864
 865                 do {
 866                         p = j ? tree_to_bkey(b, t,
 867                                         __inorder_to_eytzinger1(j--,
 868                                                         t->size - 1, t->extra))
 869                               : btree_bkey_first(b, t);
 870                 } while (p >= k);
 871                 break;
 872         case BSET_RW_AUX_TREE:
 873                 offset = __btree_node_key_to_offset(b, k);
 874                 j = rw_aux_tree_bsearch(b, t, offset);
 875                 p = j ? rw_aux_to_bkey(b, t, j - 1)
 876                       : btree_bkey_first(b, t);
 877                 break;
 878         }
 879
 880         return p;
 881 }
 882
 883 struct bkey_packed *bch2_bkey_prev_filter(struct btree *b,
 884                                           struct bset_tree *t,
 885                                           struct bkey_packed *k,
 886                                           unsigned min_key_type)
 887 {
 888         struct bkey_packed *p, *i, *ret = NULL, *orig_k = k;
 889
 890         while ((p = __bkey_prev(b, t, k)) && !ret) {
 891                 for (i = p; i != k; i = bkey_p_next(i))
 892                         if (i->type >= min_key_type)
 893                                 ret = i;
 894
 895                 k = p;
 896         }
 897
 898         if (bch2_expensive_debug_checks) {
 899                 BUG_ON(ret >= orig_k);
 900
 901                 for (i = ret
 902                         ? bkey_p_next(ret)
 903                         : btree_bkey_first(b, t);
 904                      i != orig_k;
 905                      i = bkey_p_next(i))
 906                         BUG_ON(i->type >= min_key_type);
 907         }
 908
 909         return ret;
 910 }
 911
 912 /* Insert */
 913
 914 static void bch2_bset_fix_lookup_table(struct btree *b,
 915                                        struct bset_tree *t,
 916                                        struct bkey_packed *_where,
 917                                        unsigned clobber_u64s,
 918                                        unsigned new_u64s)
 919 {
 920         int shift = new_u64s - clobber_u64s;
 921         unsigned l, j, where = __btree_node_key_to_offset(b, _where);
 922
 923         EBUG_ON(bset_has_ro_aux_tree(t));
 924
 925         if (!bset_has_rw_aux_tree(t))
 926                 return;
 927
 928         /* returns first entry >= where */
 929         l = rw_aux_tree_bsearch(b, t, where);
 930
 931         if (!l) /* never delete first entry */
 932                 l++;
 933         else if (l < t->size &&
 934                  where < t->end_offset &&
 935                  rw_aux_tree(b, t)[l].offset == where)
 936                 rw_aux_tree_set(b, t, l++, _where);
 937
 938         /* l now > where */
 939
 940         for (j = l;
 941              j < t->size &&
 942              rw_aux_tree(b, t)[j].offset < where + clobber_u64s;
 943              j++)
 944                 ;
 945
 946         if (j < t->size &&
 947             rw_aux_tree(b, t)[j].offset + shift ==
 948             rw_aux_tree(b, t)[l - 1].offset)
 949                 j++;
 950
 951         memmove(&rw_aux_tree(b, t)[l],
 952                 &rw_aux_tree(b, t)[j],
 953                 (void *) &rw_aux_tree(b, t)[t->size] -
 954                 (void *) &rw_aux_tree(b, t)[j]);
 955         t->size -= j - l;
 956
 957         for (j = l; j < t->size; j++)
 958                 rw_aux_tree(b, t)[j].offset += shift;
 959
 960         EBUG_ON(l < t->size &&
 961                 rw_aux_tree(b, t)[l].offset ==
 962                 rw_aux_tree(b, t)[l - 1].offset);
 963
 964         if (t->size < bset_rw_tree_capacity(b, t) &&
 965             (l < t->size
 966              ? rw_aux_tree(b, t)[l].offset
 967              : t->end_offset) -
 968             rw_aux_tree(b, t)[l - 1].offset >
 969             L1_CACHE_BYTES / sizeof(u64)) {
 970                 struct bkey_packed *start = rw_aux_to_bkey(b, t, l - 1);
 971                 struct bkey_packed *end = l < t->size
 972                         ? rw_aux_to_bkey(b, t, l)
 973                         : btree_bkey_last(b, t);
 974                 struct bkey_packed *k = start;
 975
 976                 while (1) {
 977                         k = bkey_p_next(k);
 978                         if (k == end)
 979                                 break;
 980
 981                         if ((void *) k - (void *) start >= L1_CACHE_BYTES) {
 982                                 memmove(&rw_aux_tree(b, t)[l + 1],
 983                                         &rw_aux_tree(b, t)[l],
 984                                         (void *) &rw_aux_tree(b, t)[t->size] -
 985                                         (void *) &rw_aux_tree(b, t)[l]);
 986                                 t->size++;
 987                                 rw_aux_tree_set(b, t, l, k);
 988                                 break;
 989                         }
 990                 }
 991         }
 992
 993         bch2_bset_verify_rw_aux_tree(b, t);
 994         bset_aux_tree_verify(b);
 995 }
 996
 997 void bch2_bset_insert(struct btree *b,
 998                       struct btree_node_iter *iter,
 999                       struct bkey_packed *where,
1000                       struct bkey_i *insert,
1001                       unsigned clobber_u64s)
1002 {
1003         struct bkey_format *f = &b->format;
1004         struct bset_tree *t = bset_tree_last(b);
1005         struct bkey_packed packed, *src = bkey_to_packed(insert);
1006
1007         bch2_bset_verify_rw_aux_tree(b, t);
1008         bch2_verify_insert_pos(b, where, bkey_to_packed(insert), clobber_u64s);
1009
1010         if (bch2_bkey_pack_key(&packed, &insert->k, f))
1011                 src = &packed;
1012
1013         if (!bkey_deleted(&insert->k))
1014                 btree_keys_account_key_add(&b->nr, t - b->set, src);
1015
1016         if (src->u64s != clobber_u64s) {
1017                 u64 *src_p = where->_data + clobber_u64s;
1018                 u64 *dst_p = where->_data + src->u64s;
1019
1020                 EBUG_ON((int) le16_to_cpu(bset(b, t)->u64s) <
1021                         (int) clobber_u64s - src->u64s);
1022
1023                 memmove_u64s(dst_p, src_p, btree_bkey_last(b, t)->_data - src_p);
1024                 le16_add_cpu(&bset(b, t)->u64s, src->u64s - clobber_u64s);
1025                 set_btree_bset_end(b, t);
1026         }
1027
1028         memcpy_u64s(where, src,
1029                     bkeyp_key_u64s(f, src));
1030         memcpy_u64s(bkeyp_val(f, where), &insert->v,
1031                     bkeyp_val_u64s(f, src));
1032
1033         if (src->u64s != clobber_u64s)
1034                 bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, src->u64s);
1035
1036         bch2_verify_btree_nr_keys(b);
1037 }
1038
1039 void bch2_bset_delete(struct btree *b,
1040                       struct bkey_packed *where,
1041                       unsigned clobber_u64s)
1042 {
1043         struct bset_tree *t = bset_tree_last(b);
1044         u64 *src_p = where->_data + clobber_u64s;
1045         u64 *dst_p = where->_data;
1046
1047         bch2_bset_verify_rw_aux_tree(b, t);
1048
1049         EBUG_ON(le16_to_cpu(bset(b, t)->u64s) < clobber_u64s);
1050
1051         memmove_u64s_down(dst_p, src_p, btree_bkey_last(b, t)->_data - src_p);
1052         le16_add_cpu(&bset(b, t)->u64s, -clobber_u64s);
1053         set_btree_bset_end(b, t);
1054
1055         bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, 0);
1056 }
1057
1058 /* Lookup */
1059
1060 __flatten
1061 static struct bkey_packed *bset_search_write_set(const struct btree *b,
1062                                 struct bset_tree *t,
1063                                 struct bpos *search)
1064 {
1065         unsigned l = 0, r = t->size;
1066
1067         while (l + 1 != r) {
1068                 unsigned m = (l + r) >> 1;
1069
1070                 if (bpos_lt(rw_aux_tree(b, t)[m].k, *search))
1071                         l = m;
1072                 else
1073                         r = m;
1074         }
1075
1076         return rw_aux_to_bkey(b, t, l);
1077 }
1078
1079 static inline void prefetch_four_cachelines(void *p)
1080 {
1081 #ifdef CONFIG_X86_64
1082         asm("prefetcht0 (-127 + 64 * 0)(%0);"
1083             "prefetcht0 (-127 + 64 * 1)(%0);"
1084             "prefetcht0 (-127 + 64 * 2)(%0);"
1085             "prefetcht0 (-127 + 64 * 3)(%0);"
1086             :
1087             : "r" (p + 127));
1088 #else
1089         prefetch(p + L1_CACHE_BYTES * 0);
1090         prefetch(p + L1_CACHE_BYTES * 1);
1091         prefetch(p + L1_CACHE_BYTES * 2);
1092         prefetch(p + L1_CACHE_BYTES * 3);
1093 #endif
1094 }
1095
1096 static inline bool bkey_mantissa_bits_dropped(const struct btree *b,
1097                                               const struct bkey_float *f,
1098                                               unsigned idx)
1099 {
1100 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
1101         unsigned key_bits_start = b->format.key_u64s * 64 - b->nr_key_bits;
1102
1103         return f->exponent > key_bits_start;
1104 #else
1105         unsigned key_bits_end = high_bit_offset + b->nr_key_bits;
1106
1107         return f->exponent + BKEY_MANTISSA_BITS < key_bits_end;
1108 #endif
1109 }
1110
1111 __flatten
1112 static struct bkey_packed *bset_search_tree(const struct btree *b,
1113                                 const struct bset_tree *t,
1114                                 const struct bpos *search,
1115                                 const struct bkey_packed *packed_search)
1116 {
1117         struct ro_aux_tree *base = ro_aux_tree_base(b, t);
1118         struct bkey_float *f;
1119         struct bkey_packed *k;
1120         unsigned inorder, n = 1, l, r;
1121         int cmp;
1122
1123         do {
1124                 if (likely(n << 4 < t->size))
1125                         prefetch(&base->f[n << 4]);
1126
1127                 f = &base->f[n];
1128                 if (unlikely(f->exponent >= BFLOAT_FAILED))
1129                         goto slowpath;
1130
1131                 l = f->mantissa;
1132                 r = bkey_mantissa(packed_search, f, n);
1133
1134                 if (unlikely(l == r) && bkey_mantissa_bits_dropped(b, f, n))
1135                         goto slowpath;
1136
1137                 n = n * 2 + (l < r);
1138                 continue;
1139 slowpath:
1140                 k = tree_to_bkey(b, t, n);
1141                 cmp = bkey_cmp_p_or_unp(b, k, packed_search, search);
1142                 if (!cmp)
1143                         return k;
1144
1145                 n = n * 2 + (cmp < 0);
1146         } while (n < t->size);
1147
1148         inorder = __eytzinger1_to_inorder(n >> 1, t->size - 1, t->extra);
1149
1150         /*
1151          * n would have been the node we recursed to - the low bit tells us if
1152          * we recursed left or recursed right.
1153          */
1154         if (likely(!(n & 1))) {
1155                 --inorder;
1156                 if (unlikely(!inorder))
1157                         return btree_bkey_first(b, t);
1158
1159                 f = &base->f[eytzinger1_prev(n >> 1, t->size - 1)];
1160         }
1161
1162         return cacheline_to_bkey(b, t, inorder, f->key_offset);
1163 }
1164
1165 static __always_inline __flatten
1166 struct bkey_packed *__bch2_bset_search(struct btree *b,
1167                                 struct bset_tree *t,
1168                                 struct bpos *search,
1169                                 const struct bkey_packed *lossy_packed_search)
1170 {
1171
1172         /*
1173          * First, we search for a cacheline, then lastly we do a linear search
1174          * within that cacheline.
1175          *
1176          * To search for the cacheline, there's three different possibilities:
1177          *  * The set is too small to have a search tree, so we just do a linear
1178          *    search over the whole set.
1179          *  * The set is the one we're currently inserting into; keeping a full
1180          *    auxiliary search tree up to date would be too expensive, so we
1181          *    use a much simpler lookup table to do a binary search -
1182          *    bset_search_write_set().
1183          *  * Or we use the auxiliary search tree we constructed earlier -
1184          *    bset_search_tree()
1185          */
1186
1187         switch (bset_aux_tree_type(t)) {
1188         case BSET_NO_AUX_TREE:
1189                 return btree_bkey_first(b, t);
1190         case BSET_RW_AUX_TREE:
1191                 return bset_search_write_set(b, t, search);
1192         case BSET_RO_AUX_TREE:
1193                 return bset_search_tree(b, t, search, lossy_packed_search);
1194         default:
1195                 unreachable();
1196         }
1197 }
1198
1199 static __always_inline __flatten
1200 struct bkey_packed *bch2_bset_search_linear(struct btree *b,
1201                                 struct bset_tree *t,
1202                                 struct bpos *search,
1203                                 struct bkey_packed *packed_search,
1204                                 const struct bkey_packed *lossy_packed_search,
1205                                 struct bkey_packed *m)
1206 {
1207         if (lossy_packed_search)
1208                 while (m != btree_bkey_last(b, t) &&
1209                        bkey_iter_cmp_p_or_unp(b, m,
1210                                         lossy_packed_search, search) < 0)
1211                         m = bkey_p_next(m);
1212
1213         if (!packed_search)
1214                 while (m != btree_bkey_last(b, t) &&
1215                        bkey_iter_pos_cmp(b, m, search) < 0)
1216                         m = bkey_p_next(m);
1217
1218         if (bch2_expensive_debug_checks) {
1219                 struct bkey_packed *prev = bch2_bkey_prev_all(b, t, m);
1220
1221                 BUG_ON(prev &&
1222                        bkey_iter_cmp_p_or_unp(b, prev,
1223                                         packed_search, search) >= 0);
1224         }
1225
1226         return m;
1227 }
1228
1229 /* Btree node iterator */
1230
1231 static inline void __bch2_btree_node_iter_push(struct btree_node_iter *iter,
1232                               struct btree *b,
1233                               const struct bkey_packed *k,
1234                               const struct bkey_packed *end)
1235 {
1236         if (k != end) {
1237                 struct btree_node_iter_set *pos;
1238
1239                 btree_node_iter_for_each(iter, pos)
1240                         ;
1241
1242                 BUG_ON(pos >= iter->data + ARRAY_SIZE(iter->data));
1243                 *pos = (struct btree_node_iter_set) {
1244                         __btree_node_key_to_offset(b, k),
1245                         __btree_node_key_to_offset(b, end)
1246                 };
1247         }
1248 }
1249
1250 void bch2_btree_node_iter_push(struct btree_node_iter *iter,
1251                                struct btree *b,
1252                                const struct bkey_packed *k,
1253                                const struct bkey_packed *end)
1254 {
1255         __bch2_btree_node_iter_push(iter, b, k, end);
1256         bch2_btree_node_iter_sort(iter, b);
1257 }
1258
1259 noinline __flatten __cold
1260 static void btree_node_iter_init_pack_failed(struct btree_node_iter *iter,
1261                               struct btree *b, struct bpos *search)
1262 {
1263         struct bkey_packed *k;
1264
1265         trace_bkey_pack_pos_fail(search);
1266
1267         bch2_btree_node_iter_init_from_start(iter, b);
1268
1269         while ((k = bch2_btree_node_iter_peek(iter, b)) &&
1270                bkey_iter_pos_cmp(b, k, search) < 0)
1271                 bch2_btree_node_iter_advance(iter, b);
1272 }
1273
1274 /**
1275  * bch_btree_node_iter_init - initialize a btree node iterator, starting from a
1276  * given position
1277  *
1278  * Main entry point to the lookup code for individual btree nodes:
1279  *
1280  * NOTE:
1281  *
1282  * When you don't filter out deleted keys, btree nodes _do_ contain duplicate
1283  * keys. This doesn't matter for most code, but it does matter for lookups.
1284  *
1285  * Some adjacent keys with a string of equal keys:
1286  *      i j k k k k l m
1287  *
1288  * If you search for k, the lookup code isn't guaranteed to return you any
1289  * specific k. The lookup code is conceptually doing a binary search and
1290  * iterating backwards is very expensive so if the pivot happens to land at the
1291  * last k that's what you'll get.
1292  *
1293  * This works out ok, but it's something to be aware of:
1294  *
1295  *  - For non extents, we guarantee that the live key comes last - see
1296  *    btree_node_iter_cmp(), keys_out_of_order(). So the duplicates you don't
1297  *    see will only be deleted keys you don't care about.
1298  *
1299  *  - For extents, deleted keys sort last (see the comment at the top of this
1300  *    file). But when you're searching for extents, you actually want the first
1301  *    key strictly greater than your search key - an extent that compares equal
1302  *    to the search key is going to have 0 sectors after the search key.
1303  *
1304  *    But this does mean that we can't just search for
1305  *    bpos_successor(start_of_range) to get the first extent that overlaps with
1306  *    the range we want - if we're unlucky and there's an extent that ends
1307  *    exactly where we searched, then there could be a deleted key at the same
1308  *    position and we'd get that when we search instead of the preceding extent
1309  *    we needed.
1310  *
1311  *    So we've got to search for start_of_range, then after the lookup iterate
1312  *    past any extents that compare equal to the position we searched for.
1313  */
1314 __flatten
1315 void bch2_btree_node_iter_init(struct btree_node_iter *iter,
1316                                struct btree *b, struct bpos *search)
1317 {
1318         struct bkey_packed p, *packed_search = NULL;
1319         struct btree_node_iter_set *pos = iter->data;
1320         struct bkey_packed *k[MAX_BSETS];
1321         unsigned i;
1322
1323         EBUG_ON(bpos_lt(*search, b->data->min_key));
1324         EBUG_ON(bpos_gt(*search, b->data->max_key));
1325         bset_aux_tree_verify(b);
1326
1327         memset(iter, 0, sizeof(*iter));
1328
1329         switch (bch2_bkey_pack_pos_lossy(&p, *search, b)) {
1330         case BKEY_PACK_POS_EXACT:
1331                 packed_search = &p;
1332                 break;
1333         case BKEY_PACK_POS_SMALLER:
1334                 packed_search = NULL;
1335                 break;
1336         case BKEY_PACK_POS_FAIL:
1337                 btree_node_iter_init_pack_failed(iter, b, search);
1338                 return;
1339         }
1340
1341         for (i = 0; i < b->nsets; i++) {
1342                 k[i] = __bch2_bset_search(b, b->set + i, search, &p);
1343                 prefetch_four_cachelines(k[i]);
1344         }
1345
1346         for (i = 0; i < b->nsets; i++) {
1347                 struct bset_tree *t = b->set + i;
1348                 struct bkey_packed *end = btree_bkey_last(b, t);
1349
1350                 k[i] = bch2_bset_search_linear(b, t, search,
1351                                                packed_search, &p, k[i]);
1352                 if (k[i] != end)
1353                         *pos++ = (struct btree_node_iter_set) {
1354                                 __btree_node_key_to_offset(b, k[i]),
1355                                 __btree_node_key_to_offset(b, end)
1356                         };
1357         }
1358
1359         bch2_btree_node_iter_sort(iter, b);
1360 }
1361
1362 void bch2_btree_node_iter_init_from_start(struct btree_node_iter *iter,
1363                                           struct btree *b)
1364 {
1365         struct bset_tree *t;
1366
1367         memset(iter, 0, sizeof(*iter));
1368
1369         for_each_bset(b, t)
1370                 __bch2_btree_node_iter_push(iter, b,
1371                                            btree_bkey_first(b, t),
1372                                            btree_bkey_last(b, t));
1373         bch2_btree_node_iter_sort(iter, b);
1374 }
1375
1376 struct bkey_packed *bch2_btree_node_iter_bset_pos(struct btree_node_iter *iter,
1377                                                   struct btree *b,
1378                                                   struct bset_tree *t)
1379 {
1380         struct btree_node_iter_set *set;
1381
1382         btree_node_iter_for_each(iter, set)
1383                 if (set->end == t->end_offset)
1384                         return __btree_node_offset_to_key(b, set->k);
1385
1386         return btree_bkey_last(b, t);
1387 }
1388
1389 static inline bool btree_node_iter_sort_two(struct btree_node_iter *iter,
1390                                             struct btree *b,
1391                                             unsigned first)
1392 {
1393         bool ret;
1394
1395         if ((ret = (btree_node_iter_cmp(b,
1396                                         iter->data[first],
1397                                         iter->data[first + 1]) > 0)))
1398                 swap(iter->data[first], iter->data[first + 1]);
1399         return ret;
1400 }
1401
1402 void bch2_btree_node_iter_sort(struct btree_node_iter *iter,
1403                                struct btree *b)
1404 {
1405         /* unrolled bubble sort: */
1406
1407         if (!__btree_node_iter_set_end(iter, 2)) {
1408                 btree_node_iter_sort_two(iter, b, 0);
1409                 btree_node_iter_sort_two(iter, b, 1);
1410         }
1411
1412         if (!__btree_node_iter_set_end(iter, 1))
1413                 btree_node_iter_sort_two(iter, b, 0);
1414 }
1415
1416 void bch2_btree_node_iter_set_drop(struct btree_node_iter *iter,
1417                                    struct btree_node_iter_set *set)
1418 {
1419         struct btree_node_iter_set *last =
1420                 iter->data + ARRAY_SIZE(iter->data) - 1;
1421
1422         memmove(&set[0], &set[1], (void *) last - (void *) set);
1423         *last = (struct btree_node_iter_set) { 0, 0 };
1424 }
1425
1426 static inline void __bch2_btree_node_iter_advance(struct btree_node_iter *iter,
1427                                                   struct btree *b)
1428 {
1429         iter->data->k += __bch2_btree_node_iter_peek_all(iter, b)->u64s;
1430
1431         EBUG_ON(iter->data->k > iter->data->end);
1432
1433         if (unlikely(__btree_node_iter_set_end(iter, 0))) {
1434                 /* avoid an expensive memmove call: */
1435                 iter->data[0] = iter->data[1];
1436                 iter->data[1] = iter->data[2];
1437                 iter->data[2] = (struct btree_node_iter_set) { 0, 0 };
1438                 return;
1439         }
1440
1441         if (__btree_node_iter_set_end(iter, 1))
1442                 return;
1443
1444         if (!btree_node_iter_sort_two(iter, b, 0))
1445                 return;
1446
1447         if (__btree_node_iter_set_end(iter, 2))
1448                 return;
1449
1450         btree_node_iter_sort_two(iter, b, 1);
1451 }
1452
1453 void bch2_btree_node_iter_advance(struct btree_node_iter *iter,
1454                                   struct btree *b)
1455 {
1456         if (bch2_expensive_debug_checks) {
1457                 bch2_btree_node_iter_verify(iter, b);
1458                 bch2_btree_node_iter_next_check(iter, b);
1459         }
1460
1461         __bch2_btree_node_iter_advance(iter, b);
1462 }
1463
1464 /*
1465  * Expensive:
1466  */
1467 struct bkey_packed *bch2_btree_node_iter_prev_all(struct btree_node_iter *iter,
1468                                                   struct btree *b)
1469 {
1470         struct bkey_packed *k, *prev = NULL;
1471         struct btree_node_iter_set *set;
1472         struct bset_tree *t;
1473         unsigned end = 0;
1474
1475         if (bch2_expensive_debug_checks)
1476                 bch2_btree_node_iter_verify(iter, b);
1477
1478         for_each_bset(b, t) {
1479                 k = bch2_bkey_prev_all(b, t,
1480                         bch2_btree_node_iter_bset_pos(iter, b, t));
1481                 if (k &&
1482                     (!prev || bkey_iter_cmp(b, k, prev) > 0)) {
1483                         prev = k;
1484                         end = t->end_offset;
1485                 }
1486         }
1487
1488         if (!prev)
1489                 return NULL;
1490
1491         /*
1492          * We're manually memmoving instead of just calling sort() to ensure the
1493          * prev we picked ends up in slot 0 - sort won't necessarily put it
1494          * there because of duplicate deleted keys:
1495          */
1496         btree_node_iter_for_each(iter, set)
1497                 if (set->end == end)
1498                         goto found;
1499
1500         BUG_ON(set != &iter->data[__btree_node_iter_used(iter)]);
1501 found:
1502         BUG_ON(set >= iter->data + ARRAY_SIZE(iter->data));
1503
1504         memmove(&iter->data[1],
1505                 &iter->data[0],
1506                 (void *) set - (void *) &iter->data[0]);
1507
1508         iter->data[0].k = __btree_node_key_to_offset(b, prev);
1509         iter->data[0].end = end;
1510
1511         if (bch2_expensive_debug_checks)
1512                 bch2_btree_node_iter_verify(iter, b);
1513         return prev;
1514 }
1515
1516 struct bkey_packed *bch2_btree_node_iter_prev(struct btree_node_iter *iter,
1517                                               struct btree *b)
1518 {
1519         struct bkey_packed *prev;
1520
1521         do {
1522                 prev = bch2_btree_node_iter_prev_all(iter, b);
1523         } while (prev && bkey_deleted(prev));
1524
1525         return prev;
1526 }
1527
1528 struct bkey_s_c bch2_btree_node_iter_peek_unpack(struct btree_node_iter *iter,
1529                                                  struct btree *b,
1530                                                  struct bkey *u)
1531 {
1532         struct bkey_packed *k = bch2_btree_node_iter_peek(iter, b);
1533
1534         return k ? bkey_disassemble(b, k, u) : bkey_s_c_null;
1535 }
1536
1537 /* Mergesort */
1538
1539 void bch2_btree_keys_stats(const struct btree *b, struct bset_stats *stats)
1540 {
1541         const struct bset_tree *t;
1542
1543         for_each_bset(b, t) {
1544                 enum bset_aux_tree_type type = bset_aux_tree_type(t);
1545                 size_t j;
1546
1547                 stats->sets[type].nr++;
1548                 stats->sets[type].bytes += le16_to_cpu(bset(b, t)->u64s) *
1549                         sizeof(u64);
1550
1551                 if (bset_has_ro_aux_tree(t)) {
1552                         stats->floats += t->size - 1;
1553
1554                         for (j = 1; j < t->size; j++)
1555                                 stats->failed +=
1556                                         bkey_float(b, t, j)->exponent ==
1557                                         BFLOAT_FAILED;
1558                 }
1559         }
1560 }
1561
1562 void bch2_bfloat_to_text(struct printbuf *out, struct btree *b,
1563                          struct bkey_packed *k)
1564 {
1565         struct bset_tree *t = bch2_bkey_to_bset(b, k);
1566         struct bkey uk;
1567         unsigned j, inorder;
1568
1569         if (!bset_has_ro_aux_tree(t))
1570                 return;
1571
1572         inorder = bkey_to_cacheline(b, t, k);
1573         if (!inorder || inorder >= t->size)
1574                 return;
1575
1576         j = __inorder_to_eytzinger1(inorder, t->size - 1, t->extra);
1577         if (k != tree_to_bkey(b, t, j))
1578                 return;
1579
1580         switch (bkey_float(b, t, j)->exponent) {
1581         case BFLOAT_FAILED:
1582                 uk = bkey_unpack_key(b, k);
1583                 prt_printf(out,
1584                        "    failed unpacked at depth %u\n"
1585                        "\t",
1586                        ilog2(j));
1587                 bch2_bpos_to_text(out, uk.p);
1588                 prt_printf(out, "\n");
1589                 break;
1590         }
1591 }