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 "trace.h"
  14 #include "util.h"
  15
  16 #include <asm/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                 k = bkey_disassemble(b, _k, &uk);
  72
  73                 printbuf_reset(&buf);
  74                 if (c)
  75                         bch2_bkey_val_to_text(&buf, c, k);
  76                 else
  77                         bch2_bkey_to_text(&buf, k.k);
  78                 printk(KERN_ERR "block %u key %5zu: %s\n", set,
  79                        _k->_data - i->_data, buf.buf);
  80
  81                 if (_n == vstruct_last(i))
  82                         continue;
  83
  84                 n = bkey_unpack_key(b, _n);
  85
  86                 if (bpos_lt(n.p, k.k->p)) {
  87                         printk(KERN_ERR "Key skipped backwards\n");
  88                         continue;
  89                 }
  90
  91                 if (!bkey_deleted(k.k) && bpos_eq(n.p, k.k->p))
  92                         printk(KERN_ERR "Duplicate keys\n");
  93         }
  94
  95         printbuf_exit(&buf);
  96 }
  97
  98 void bch2_dump_btree_node(struct bch_fs *c, struct btree *b)
  99 {
 100         struct bset_tree *t;
 101
 102         console_lock();
 103         for_each_bset(b, t)
 104                 bch2_dump_bset(c, b, bset(b, t), t - b->set);
 105         console_unlock();
 106 }
 107
 108 void bch2_dump_btree_node_iter(struct btree *b,
 109                               struct btree_node_iter *iter)
 110 {
 111         struct btree_node_iter_set *set;
 112         struct printbuf buf = PRINTBUF;
 113
 114         printk(KERN_ERR "btree node iter with %u/%u sets:\n",
 115                __btree_node_iter_used(iter), b->nsets);
 116
 117         btree_node_iter_for_each(iter, set) {
 118                 struct bkey_packed *k = __btree_node_offset_to_key(b, set->k);
 119                 struct bset_tree *t = bch2_bkey_to_bset(b, k);
 120                 struct bkey uk = bkey_unpack_key(b, k);
 121
 122                 printbuf_reset(&buf);
 123                 bch2_bkey_to_text(&buf, &uk);
 124                 printk(KERN_ERR "set %zu key %u: %s\n",
 125                        t - b->set, set->k, buf.buf);
 126         }
 127
 128         printbuf_exit(&buf);
 129 }
 130
 131 #ifdef CONFIG_BCACHEFS_DEBUG
 132
 133 void __bch2_verify_btree_nr_keys(struct btree *b)
 134 {
 135         struct bset_tree *t;
 136         struct bkey_packed *k;
 137         struct btree_nr_keys nr = { 0 };
 138
 139         for_each_bset(b, t)
 140                 bset_tree_for_each_key(b, t, k)
 141                         if (!bkey_deleted(k))
 142                                 btree_keys_account_key_add(&nr, t - b->set, k);
 143
 144         BUG_ON(memcmp(&nr, &b->nr, sizeof(nr)));
 145 }
 146
 147 static void bch2_btree_node_iter_next_check(struct btree_node_iter *_iter,
 148                                             struct btree *b)
 149 {
 150         struct btree_node_iter iter = *_iter;
 151         const struct bkey_packed *k, *n;
 152
 153         k = bch2_btree_node_iter_peek_all(&iter, b);
 154         __bch2_btree_node_iter_advance(&iter, b);
 155         n = bch2_btree_node_iter_peek_all(&iter, b);
 156
 157         bkey_unpack_key(b, k);
 158
 159         if (n &&
 160             bkey_iter_cmp(b, k, n) > 0) {
 161                 struct btree_node_iter_set *set;
 162                 struct bkey ku = bkey_unpack_key(b, k);
 163                 struct bkey nu = bkey_unpack_key(b, n);
 164                 struct printbuf buf1 = PRINTBUF;
 165                 struct printbuf buf2 = PRINTBUF;
 166
 167                 bch2_dump_btree_node(NULL, b);
 168                 bch2_bkey_to_text(&buf1, &ku);
 169                 bch2_bkey_to_text(&buf2, &nu);
 170                 printk(KERN_ERR "out of order/overlapping:\n%s\n%s\n",
 171                        buf1.buf, buf2.buf);
 172                 printk(KERN_ERR "iter was:");
 173
 174                 btree_node_iter_for_each(_iter, set) {
 175                         struct bkey_packed *k = __btree_node_offset_to_key(b, set->k);
 176                         struct bset_tree *t = bch2_bkey_to_bset(b, k);
 177                         printk(" [%zi %zi]", t - b->set,
 178                                k->_data - bset(b, t)->_data);
 179                 }
 180                 panic("\n");
 181         }
 182 }
 183
 184 void bch2_btree_node_iter_verify(struct btree_node_iter *iter,
 185                                  struct btree *b)
 186 {
 187         struct btree_node_iter_set *set, *s2;
 188         struct bkey_packed *k, *p;
 189         struct bset_tree *t;
 190
 191         if (bch2_btree_node_iter_end(iter))
 192                 return;
 193
 194         /* Verify no duplicates: */
 195         btree_node_iter_for_each(iter, set) {
 196                 BUG_ON(set->k > set->end);
 197                 btree_node_iter_for_each(iter, s2)
 198                         BUG_ON(set != s2 && set->end == s2->end);
 199         }
 200
 201         /* Verify that set->end is correct: */
 202         btree_node_iter_for_each(iter, set) {
 203                 for_each_bset(b, t)
 204                         if (set->end == t->end_offset)
 205                                 goto found;
 206                 BUG();
 207 found:
 208                 BUG_ON(set->k < btree_bkey_first_offset(t) ||
 209                        set->k >= t->end_offset);
 210         }
 211
 212         /* Verify iterator is sorted: */
 213         btree_node_iter_for_each(iter, set)
 214                 BUG_ON(set != iter->data &&
 215                        btree_node_iter_cmp(b, set[-1], set[0]) > 0);
 216
 217         k = bch2_btree_node_iter_peek_all(iter, b);
 218
 219         for_each_bset(b, t) {
 220                 if (iter->data[0].end == t->end_offset)
 221                         continue;
 222
 223                 p = bch2_bkey_prev_all(b, t,
 224                         bch2_btree_node_iter_bset_pos(iter, b, t));
 225
 226                 BUG_ON(p && bkey_iter_cmp(b, k, p) < 0);
 227         }
 228 }
 229
 230 void bch2_verify_insert_pos(struct btree *b, struct bkey_packed *where,
 231                             struct bkey_packed *insert, unsigned clobber_u64s)
 232 {
 233         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 234         struct bkey_packed *prev = bch2_bkey_prev_all(b, t, where);
 235         struct bkey_packed *next = (void *) (where->_data + clobber_u64s);
 236         struct printbuf buf1 = PRINTBUF;
 237         struct printbuf buf2 = PRINTBUF;
 238 #if 0
 239         BUG_ON(prev &&
 240                bkey_iter_cmp(b, prev, insert) > 0);
 241 #else
 242         if (prev &&
 243             bkey_iter_cmp(b, prev, insert) > 0) {
 244                 struct bkey k1 = bkey_unpack_key(b, prev);
 245                 struct bkey k2 = bkey_unpack_key(b, insert);
 246
 247                 bch2_dump_btree_node(NULL, b);
 248                 bch2_bkey_to_text(&buf1, &k1);
 249                 bch2_bkey_to_text(&buf2, &k2);
 250
 251                 panic("prev > insert:\n"
 252                       "prev    key %s\n"
 253                       "insert  key %s\n",
 254                       buf1.buf, buf2.buf);
 255         }
 256 #endif
 257 #if 0
 258         BUG_ON(next != btree_bkey_last(b, t) &&
 259                bkey_iter_cmp(b, insert, next) > 0);
 260 #else
 261         if (next != btree_bkey_last(b, t) &&
 262             bkey_iter_cmp(b, insert, next) > 0) {
 263                 struct bkey k1 = bkey_unpack_key(b, insert);
 264                 struct bkey k2 = bkey_unpack_key(b, next);
 265
 266                 bch2_dump_btree_node(NULL, b);
 267                 bch2_bkey_to_text(&buf1, &k1);
 268                 bch2_bkey_to_text(&buf2, &k2);
 269
 270                 panic("insert > next:\n"
 271                       "insert  key %s\n"
 272                       "next    key %s\n",
 273                       buf1.buf, buf2.buf);
 274         }
 275 #endif
 276 }
 277
 278 #else
 279
 280 static inline void bch2_btree_node_iter_next_check(struct btree_node_iter *iter,
 281                                                    struct btree *b) {}
 282
 283 #endif
 284
 285 /* Auxiliary search trees */
 286
 287 #define BFLOAT_FAILED_UNPACKED  U8_MAX
 288 #define BFLOAT_FAILED           U8_MAX
 289
 290 struct bkey_float {
 291         u8              exponent;
 292         u8              key_offset;
 293         u16             mantissa;
 294 };
 295 #define BKEY_MANTISSA_BITS      16
 296
 297 static unsigned bkey_float_byte_offset(unsigned idx)
 298 {
 299         return idx * sizeof(struct bkey_float);
 300 }
 301
 302 struct ro_aux_tree {
 303         struct bkey_float       f[0];
 304 };
 305
 306 struct rw_aux_tree {
 307         u16             offset;
 308         struct bpos     k;
 309 };
 310
 311 static unsigned bset_aux_tree_buf_end(const struct bset_tree *t)
 312 {
 313         BUG_ON(t->aux_data_offset == U16_MAX);
 314
 315         switch (bset_aux_tree_type(t)) {
 316         case BSET_NO_AUX_TREE:
 317                 return t->aux_data_offset;
 318         case BSET_RO_AUX_TREE:
 319                 return t->aux_data_offset +
 320                         DIV_ROUND_UP(t->size * sizeof(struct bkey_float) +
 321                                      t->size * sizeof(u8), 8);
 322         case BSET_RW_AUX_TREE:
 323                 return t->aux_data_offset +
 324                         DIV_ROUND_UP(sizeof(struct rw_aux_tree) * t->size, 8);
 325         default:
 326                 BUG();
 327         }
 328 }
 329
 330 static unsigned bset_aux_tree_buf_start(const struct btree *b,
 331                                         const struct bset_tree *t)
 332 {
 333         return t == b->set
 334                 ? DIV_ROUND_UP(b->unpack_fn_len, 8)
 335                 : bset_aux_tree_buf_end(t - 1);
 336 }
 337
 338 static void *__aux_tree_base(const struct btree *b,
 339                              const struct bset_tree *t)
 340 {
 341         return b->aux_data + t->aux_data_offset * 8;
 342 }
 343
 344 static struct ro_aux_tree *ro_aux_tree_base(const struct btree *b,
 345                                             const struct bset_tree *t)
 346 {
 347         EBUG_ON(bset_aux_tree_type(t) != BSET_RO_AUX_TREE);
 348
 349         return __aux_tree_base(b, t);
 350 }
 351
 352 static u8 *ro_aux_tree_prev(const struct btree *b,
 353                             const struct bset_tree *t)
 354 {
 355         EBUG_ON(bset_aux_tree_type(t) != BSET_RO_AUX_TREE);
 356
 357         return __aux_tree_base(b, t) + bkey_float_byte_offset(t->size);
 358 }
 359
 360 static struct bkey_float *bkey_float(const struct btree *b,
 361                                      const struct bset_tree *t,
 362                                      unsigned idx)
 363 {
 364         return ro_aux_tree_base(b, t)->f + idx;
 365 }
 366
 367 static void bset_aux_tree_verify(const struct btree *b)
 368 {
 369 #ifdef CONFIG_BCACHEFS_DEBUG
 370         const struct bset_tree *t;
 371
 372         for_each_bset(b, t) {
 373                 if (t->aux_data_offset == U16_MAX)
 374                         continue;
 375
 376                 BUG_ON(t != b->set &&
 377                        t[-1].aux_data_offset == U16_MAX);
 378
 379                 BUG_ON(t->aux_data_offset < bset_aux_tree_buf_start(b, t));
 380                 BUG_ON(t->aux_data_offset > btree_aux_data_u64s(b));
 381                 BUG_ON(bset_aux_tree_buf_end(t) > btree_aux_data_u64s(b));
 382         }
 383 #endif
 384 }
 385
 386 void bch2_btree_keys_init(struct btree *b)
 387 {
 388         unsigned i;
 389
 390         b->nsets                = 0;
 391         memset(&b->nr, 0, sizeof(b->nr));
 392
 393         for (i = 0; i < MAX_BSETS; i++)
 394                 b->set[i].data_offset = U16_MAX;
 395
 396         bch2_bset_set_no_aux_tree(b, b->set);
 397 }
 398
 399 /* Binary tree stuff for auxiliary search trees */
 400
 401 /*
 402  * Cacheline/offset <-> bkey pointer arithmetic:
 403  *
 404  * t->tree is a binary search tree in an array; each node corresponds to a key
 405  * in one cacheline in t->set (BSET_CACHELINE bytes).
 406  *
 407  * This means we don't have to store the full index of the key that a node in
 408  * the binary tree points to; eytzinger1_to_inorder() gives us the cacheline, and
 409  * then bkey_float->m gives us the offset within that cacheline, in units of 8
 410  * bytes.
 411  *
 412  * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
 413  * make this work.
 414  *
 415  * To construct the bfloat for an arbitrary key we need to know what the key
 416  * immediately preceding it is: we have to check if the two keys differ in the
 417  * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
 418  * of the previous key so we can walk backwards to it from t->tree[j]'s key.
 419  */
 420
 421 static inline void *bset_cacheline(const struct btree *b,
 422                                    const struct bset_tree *t,
 423                                    unsigned cacheline)
 424 {
 425         return (void *) round_down((unsigned long) btree_bkey_first(b, t),
 426                                    L1_CACHE_BYTES) +
 427                 cacheline * BSET_CACHELINE;
 428 }
 429
 430 static struct bkey_packed *cacheline_to_bkey(const struct btree *b,
 431                                              const struct bset_tree *t,
 432                                              unsigned cacheline,
 433                                              unsigned offset)
 434 {
 435         return bset_cacheline(b, t, cacheline) + offset * 8;
 436 }
 437
 438 static unsigned bkey_to_cacheline(const struct btree *b,
 439                                   const struct bset_tree *t,
 440                                   const struct bkey_packed *k)
 441 {
 442         return ((void *) k - bset_cacheline(b, t, 0)) / BSET_CACHELINE;
 443 }
 444
 445 static ssize_t __bkey_to_cacheline_offset(const struct btree *b,
 446                                           const struct bset_tree *t,
 447                                           unsigned cacheline,
 448                                           const struct bkey_packed *k)
 449 {
 450         return (u64 *) k - (u64 *) bset_cacheline(b, t, cacheline);
 451 }
 452
 453 static unsigned bkey_to_cacheline_offset(const struct btree *b,
 454                                          const struct bset_tree *t,
 455                                          unsigned cacheline,
 456                                          const struct bkey_packed *k)
 457 {
 458         size_t m = __bkey_to_cacheline_offset(b, t, cacheline, k);
 459
 460         EBUG_ON(m > U8_MAX);
 461         return m;
 462 }
 463
 464 static inline struct bkey_packed *tree_to_bkey(const struct btree *b,
 465                                                const struct bset_tree *t,
 466                                                unsigned j)
 467 {
 468         return cacheline_to_bkey(b, t,
 469                         __eytzinger1_to_inorder(j, t->size - 1, t->extra),
 470                         bkey_float(b, t, j)->key_offset);
 471 }
 472
 473 static struct bkey_packed *tree_to_prev_bkey(const struct btree *b,
 474                                              const struct bset_tree *t,
 475                                              unsigned j)
 476 {
 477         unsigned prev_u64s = ro_aux_tree_prev(b, t)[j];
 478
 479         return (void *) (tree_to_bkey(b, t, j)->_data - prev_u64s);
 480 }
 481
 482 static struct rw_aux_tree *rw_aux_tree(const struct btree *b,
 483                                        const struct bset_tree *t)
 484 {
 485         EBUG_ON(bset_aux_tree_type(t) != BSET_RW_AUX_TREE);
 486
 487         return __aux_tree_base(b, t);
 488 }
 489
 490 /*
 491  * For the write set - the one we're currently inserting keys into - we don't
 492  * maintain a full search tree, we just keep a simple lookup table in t->prev.
 493  */
 494 static struct bkey_packed *rw_aux_to_bkey(const struct btree *b,
 495                                           struct bset_tree *t,
 496                                           unsigned j)
 497 {
 498         return __btree_node_offset_to_key(b, rw_aux_tree(b, t)[j].offset);
 499 }
 500
 501 static void rw_aux_tree_set(const struct btree *b, struct bset_tree *t,
 502                             unsigned j, struct bkey_packed *k)
 503 {
 504         EBUG_ON(k >= btree_bkey_last(b, t));
 505
 506         rw_aux_tree(b, t)[j] = (struct rw_aux_tree) {
 507                 .offset = __btree_node_key_to_offset(b, k),
 508                 .k      = bkey_unpack_pos(b, k),
 509         };
 510 }
 511
 512 static void bch2_bset_verify_rw_aux_tree(struct btree *b,
 513                                         struct bset_tree *t)
 514 {
 515         struct bkey_packed *k = btree_bkey_first(b, t);
 516         unsigned j = 0;
 517
 518         if (!bch2_expensive_debug_checks)
 519                 return;
 520
 521         BUG_ON(bset_has_ro_aux_tree(t));
 522
 523         if (!bset_has_rw_aux_tree(t))
 524                 return;
 525
 526         BUG_ON(t->size < 1);
 527         BUG_ON(rw_aux_to_bkey(b, t, j) != k);
 528
 529         goto start;
 530         while (1) {
 531                 if (rw_aux_to_bkey(b, t, j) == k) {
 532                         BUG_ON(!bpos_eq(rw_aux_tree(b, t)[j].k,
 533                                         bkey_unpack_pos(b, k)));
 534 start:
 535                         if (++j == t->size)
 536                                 break;
 537
 538                         BUG_ON(rw_aux_tree(b, t)[j].offset <=
 539                                rw_aux_tree(b, t)[j - 1].offset);
 540                 }
 541
 542                 k = bkey_p_next(k);
 543                 BUG_ON(k >= btree_bkey_last(b, t));
 544         }
 545 }
 546
 547 /* returns idx of first entry >= offset: */
 548 static unsigned rw_aux_tree_bsearch(struct btree *b,
 549                                     struct bset_tree *t,
 550                                     unsigned offset)
 551 {
 552         unsigned bset_offs = offset - btree_bkey_first_offset(t);
 553         unsigned bset_u64s = t->end_offset - btree_bkey_first_offset(t);
 554         unsigned idx = bset_u64s ? bset_offs * t->size / bset_u64s : 0;
 555
 556         EBUG_ON(bset_aux_tree_type(t) != BSET_RW_AUX_TREE);
 557         EBUG_ON(!t->size);
 558         EBUG_ON(idx > t->size);
 559
 560         while (idx < t->size &&
 561                rw_aux_tree(b, t)[idx].offset < offset)
 562                 idx++;
 563
 564         while (idx &&
 565                rw_aux_tree(b, t)[idx - 1].offset >= offset)
 566                 idx--;
 567
 568         EBUG_ON(idx < t->size &&
 569                 rw_aux_tree(b, t)[idx].offset < offset);
 570         EBUG_ON(idx && rw_aux_tree(b, t)[idx - 1].offset >= offset);
 571         EBUG_ON(idx + 1 < t->size &&
 572                 rw_aux_tree(b, t)[idx].offset ==
 573                 rw_aux_tree(b, t)[idx + 1].offset);
 574
 575         return idx;
 576 }
 577
 578 static inline unsigned bkey_mantissa(const struct bkey_packed *k,
 579                                      const struct bkey_float *f,
 580                                      unsigned idx)
 581 {
 582         u64 v;
 583
 584         EBUG_ON(!bkey_packed(k));
 585
 586         v = get_unaligned((u64 *) (((u8 *) k->_data) + (f->exponent >> 3)));
 587
 588         /*
 589          * In little endian, we're shifting off low bits (and then the bits we
 590          * want are at the low end), in big endian we're shifting off high bits
 591          * (and then the bits we want are at the high end, so we shift them
 592          * back down):
 593          */
 594 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 595         v >>= f->exponent & 7;
 596 #else
 597         v >>= 64 - (f->exponent & 7) - BKEY_MANTISSA_BITS;
 598 #endif
 599         return (u16) v;
 600 }
 601
 602 static __always_inline void make_bfloat(struct btree *b, struct bset_tree *t,
 603                                         unsigned j,
 604                                         struct bkey_packed *min_key,
 605                                         struct bkey_packed *max_key)
 606 {
 607         struct bkey_float *f = bkey_float(b, t, j);
 608         struct bkey_packed *m = tree_to_bkey(b, t, j);
 609         struct bkey_packed *l = is_power_of_2(j)
 610                 ? min_key
 611                 : tree_to_prev_bkey(b, t, j >> ffs(j));
 612         struct bkey_packed *r = is_power_of_2(j + 1)
 613                 ? max_key
 614                 : tree_to_bkey(b, t, j >> (ffz(j) + 1));
 615         unsigned mantissa;
 616         int shift, exponent, high_bit;
 617
 618         /*
 619          * for failed bfloats, the lookup code falls back to comparing against
 620          * the original key.
 621          */
 622
 623         if (!bkey_packed(l) || !bkey_packed(r) || !bkey_packed(m) ||
 624             !b->nr_key_bits) {
 625                 f->exponent = BFLOAT_FAILED_UNPACKED;
 626                 return;
 627         }
 628
 629         /*
 630          * The greatest differing bit of l and r is the first bit we must
 631          * include in the bfloat mantissa we're creating in order to do
 632          * comparisons - that bit always becomes the high bit of
 633          * bfloat->mantissa, and thus the exponent we're calculating here is
 634          * the position of what will become the low bit in bfloat->mantissa:
 635          *
 636          * Note that this may be negative - we may be running off the low end
 637          * of the key: we handle this later:
 638          */
 639         high_bit = max(bch2_bkey_greatest_differing_bit(b, l, r),
 640                        min_t(unsigned, BKEY_MANTISSA_BITS, b->nr_key_bits) - 1);
 641         exponent = high_bit - (BKEY_MANTISSA_BITS - 1);
 642
 643         /*
 644          * Then we calculate the actual shift value, from the start of the key
 645          * (k->_data), to get the key bits starting at exponent:
 646          */
 647 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 648         shift = (int) (b->format.key_u64s * 64 - b->nr_key_bits) + exponent;
 649
 650         EBUG_ON(shift + BKEY_MANTISSA_BITS > b->format.key_u64s * 64);
 651 #else
 652         shift = high_bit_offset +
 653                 b->nr_key_bits -
 654                 exponent -
 655                 BKEY_MANTISSA_BITS;
 656
 657         EBUG_ON(shift < KEY_PACKED_BITS_START);
 658 #endif
 659         EBUG_ON(shift < 0 || shift >= BFLOAT_FAILED);
 660
 661         f->exponent = shift;
 662         mantissa = bkey_mantissa(m, f, j);
 663
 664         /*
 665          * If we've got garbage bits, set them to all 1s - it's legal for the
 666          * bfloat to compare larger than the original key, but not smaller:
 667          */
 668         if (exponent < 0)
 669                 mantissa |= ~(~0U << -exponent);
 670
 671         f->mantissa = mantissa;
 672 }
 673
 674 /* bytes remaining - only valid for last bset: */
 675 static unsigned __bset_tree_capacity(const struct btree *b, const struct bset_tree *t)
 676 {
 677         bset_aux_tree_verify(b);
 678
 679         return btree_aux_data_bytes(b) - t->aux_data_offset * sizeof(u64);
 680 }
 681
 682 static unsigned bset_ro_tree_capacity(const struct btree *b, const struct bset_tree *t)
 683 {
 684         return __bset_tree_capacity(b, t) /
 685                 (sizeof(struct bkey_float) + sizeof(u8));
 686 }
 687
 688 static unsigned bset_rw_tree_capacity(const struct btree *b, const struct bset_tree *t)
 689 {
 690         return __bset_tree_capacity(b, t) / sizeof(struct rw_aux_tree);
 691 }
 692
 693 static noinline void __build_rw_aux_tree(struct btree *b, struct bset_tree *t)
 694 {
 695         struct bkey_packed *k;
 696
 697         t->size = 1;
 698         t->extra = BSET_RW_AUX_TREE_VAL;
 699         rw_aux_tree(b, t)[0].offset =
 700                 __btree_node_key_to_offset(b, btree_bkey_first(b, t));
 701
 702         bset_tree_for_each_key(b, t, k) {
 703                 if (t->size == bset_rw_tree_capacity(b, t))
 704                         break;
 705
 706                 if ((void *) k - (void *) rw_aux_to_bkey(b, t, t->size - 1) >
 707                     L1_CACHE_BYTES)
 708                         rw_aux_tree_set(b, t, t->size++, k);
 709         }
 710 }
 711
 712 static noinline void __build_ro_aux_tree(struct btree *b, struct bset_tree *t)
 713 {
 714         struct bkey_packed *prev = NULL, *k = btree_bkey_first(b, t);
 715         struct bkey_i min_key, max_key;
 716         unsigned j, cacheline = 1;
 717
 718         t->size = min(bkey_to_cacheline(b, t, btree_bkey_last(b, t)),
 719                       bset_ro_tree_capacity(b, t));
 720 retry:
 721         if (t->size < 2) {
 722                 t->size = 0;
 723                 t->extra = BSET_NO_AUX_TREE_VAL;
 724                 return;
 725         }
 726
 727         t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
 728
 729         /* First we figure out where the first key in each cacheline is */
 730         eytzinger1_for_each(j, t->size - 1) {
 731                 while (bkey_to_cacheline(b, t, k) < cacheline)
 732                         prev = k, k = bkey_p_next(k);
 733
 734                 if (k >= btree_bkey_last(b, t)) {
 735                         /* XXX: this path sucks */
 736                         t->size--;
 737                         goto retry;
 738                 }
 739
 740                 ro_aux_tree_prev(b, t)[j] = prev->u64s;
 741                 bkey_float(b, t, j)->key_offset =
 742                         bkey_to_cacheline_offset(b, t, cacheline++, k);
 743
 744                 EBUG_ON(tree_to_prev_bkey(b, t, j) != prev);
 745                 EBUG_ON(tree_to_bkey(b, t, j) != k);
 746         }
 747
 748         while (k != btree_bkey_last(b, t))
 749                 prev = k, k = bkey_p_next(k);
 750
 751         if (!bkey_pack_pos(bkey_to_packed(&min_key), b->data->min_key, b)) {
 752                 bkey_init(&min_key.k);
 753                 min_key.k.p = b->data->min_key;
 754         }
 755
 756         if (!bkey_pack_pos(bkey_to_packed(&max_key), b->data->max_key, b)) {
 757                 bkey_init(&max_key.k);
 758                 max_key.k.p = b->data->max_key;
 759         }
 760
 761         /* Then we build the tree */
 762         eytzinger1_for_each(j, t->size - 1)
 763                 make_bfloat(b, t, j,
 764                             bkey_to_packed(&min_key),
 765                             bkey_to_packed(&max_key));
 766 }
 767
 768 static void bset_alloc_tree(struct btree *b, struct bset_tree *t)
 769 {
 770         struct bset_tree *i;
 771
 772         for (i = b->set; i != t; i++)
 773                 BUG_ON(bset_has_rw_aux_tree(i));
 774
 775         bch2_bset_set_no_aux_tree(b, t);
 776
 777         /* round up to next cacheline: */
 778         t->aux_data_offset = round_up(bset_aux_tree_buf_start(b, t),
 779                                       SMP_CACHE_BYTES / sizeof(u64));
 780
 781         bset_aux_tree_verify(b);
 782 }
 783
 784 void bch2_bset_build_aux_tree(struct btree *b, struct bset_tree *t,
 785                              bool writeable)
 786 {
 787         if (writeable
 788             ? bset_has_rw_aux_tree(t)
 789             : bset_has_ro_aux_tree(t))
 790                 return;
 791
 792         bset_alloc_tree(b, t);
 793
 794         if (!__bset_tree_capacity(b, t))
 795                 return;
 796
 797         if (writeable)
 798                 __build_rw_aux_tree(b, t);
 799         else
 800                 __build_ro_aux_tree(b, t);
 801
 802         bset_aux_tree_verify(b);
 803 }
 804
 805 void bch2_bset_init_first(struct btree *b, struct bset *i)
 806 {
 807         struct bset_tree *t;
 808
 809         BUG_ON(b->nsets);
 810
 811         memset(i, 0, sizeof(*i));
 812         get_random_bytes(&i->seq, sizeof(i->seq));
 813         SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 814
 815         t = &b->set[b->nsets++];
 816         set_btree_bset(b, t, i);
 817 }
 818
 819 void bch2_bset_init_next(struct bch_fs *c, struct btree *b,
 820                          struct btree_node_entry *bne)
 821 {
 822         struct bset *i = &bne->keys;
 823         struct bset_tree *t;
 824
 825         BUG_ON(bset_byte_offset(b, bne) >= btree_bytes(c));
 826         BUG_ON((void *) bne < (void *) btree_bkey_last(b, bset_tree_last(b)));
 827         BUG_ON(b->nsets >= MAX_BSETS);
 828
 829         memset(i, 0, sizeof(*i));
 830         i->seq = btree_bset_first(b)->seq;
 831         SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 832
 833         t = &b->set[b->nsets++];
 834         set_btree_bset(b, t, i);
 835 }
 836
 837 /*
 838  * find _some_ key in the same bset as @k that precedes @k - not necessarily the
 839  * immediate predecessor:
 840  */
 841 static struct bkey_packed *__bkey_prev(struct btree *b, struct bset_tree *t,
 842                                        struct bkey_packed *k)
 843 {
 844         struct bkey_packed *p;
 845         unsigned offset;
 846         int j;
 847
 848         EBUG_ON(k < btree_bkey_first(b, t) ||
 849                 k > btree_bkey_last(b, t));
 850
 851         if (k == btree_bkey_first(b, t))
 852                 return NULL;
 853
 854         switch (bset_aux_tree_type(t)) {
 855         case BSET_NO_AUX_TREE:
 856                 p = btree_bkey_first(b, t);
 857                 break;
 858         case BSET_RO_AUX_TREE:
 859                 j = min_t(unsigned, t->size - 1, bkey_to_cacheline(b, t, k));
 860
 861                 do {
 862                         p = j ? tree_to_bkey(b, t,
 863                                         __inorder_to_eytzinger1(j--,
 864                                                         t->size - 1, t->extra))
 865                               : btree_bkey_first(b, t);
 866                 } while (p >= k);
 867                 break;
 868         case BSET_RW_AUX_TREE:
 869                 offset = __btree_node_key_to_offset(b, k);
 870                 j = rw_aux_tree_bsearch(b, t, offset);
 871                 p = j ? rw_aux_to_bkey(b, t, j - 1)
 872                       : btree_bkey_first(b, t);
 873                 break;
 874         }
 875
 876         return p;
 877 }
 878
 879 struct bkey_packed *bch2_bkey_prev_filter(struct btree *b,
 880                                           struct bset_tree *t,
 881                                           struct bkey_packed *k,
 882                                           unsigned min_key_type)
 883 {
 884         struct bkey_packed *p, *i, *ret = NULL, *orig_k = k;
 885
 886         while ((p = __bkey_prev(b, t, k)) && !ret) {
 887                 for (i = p; i != k; i = bkey_p_next(i))
 888                         if (i->type >= min_key_type)
 889                                 ret = i;
 890
 891                 k = p;
 892         }
 893
 894         if (bch2_expensive_debug_checks) {
 895                 BUG_ON(ret >= orig_k);
 896
 897                 for (i = ret
 898                         ? bkey_p_next(ret)
 899                         : btree_bkey_first(b, t);
 900                      i != orig_k;
 901                      i = bkey_p_next(i))
 902                         BUG_ON(i->type >= min_key_type);
 903         }
 904
 905         return ret;
 906 }
 907
 908 /* Insert */
 909
 910 static void bch2_bset_fix_lookup_table(struct btree *b,
 911                                        struct bset_tree *t,
 912                                        struct bkey_packed *_where,
 913                                        unsigned clobber_u64s,
 914                                        unsigned new_u64s)
 915 {
 916         int shift = new_u64s - clobber_u64s;
 917         unsigned l, j, where = __btree_node_key_to_offset(b, _where);
 918
 919         EBUG_ON(bset_has_ro_aux_tree(t));
 920
 921         if (!bset_has_rw_aux_tree(t))
 922                 return;
 923
 924         /* returns first entry >= where */
 925         l = rw_aux_tree_bsearch(b, t, where);
 926
 927         if (!l) /* never delete first entry */
 928                 l++;
 929         else if (l < t->size &&
 930                  where < t->end_offset &&
 931                  rw_aux_tree(b, t)[l].offset == where)
 932                 rw_aux_tree_set(b, t, l++, _where);
 933
 934         /* l now > where */
 935
 936         for (j = l;
 937              j < t->size &&
 938              rw_aux_tree(b, t)[j].offset < where + clobber_u64s;
 939              j++)
 940                 ;
 941
 942         if (j < t->size &&
 943             rw_aux_tree(b, t)[j].offset + shift ==
 944             rw_aux_tree(b, t)[l - 1].offset)
 945                 j++;
 946
 947         memmove(&rw_aux_tree(b, t)[l],
 948                 &rw_aux_tree(b, t)[j],
 949                 (void *) &rw_aux_tree(b, t)[t->size] -
 950                 (void *) &rw_aux_tree(b, t)[j]);
 951         t->size -= j - l;
 952
 953         for (j = l; j < t->size; j++)
 954                 rw_aux_tree(b, t)[j].offset += shift;
 955
 956         EBUG_ON(l < t->size &&
 957                 rw_aux_tree(b, t)[l].offset ==
 958                 rw_aux_tree(b, t)[l - 1].offset);
 959
 960         if (t->size < bset_rw_tree_capacity(b, t) &&
 961             (l < t->size
 962              ? rw_aux_tree(b, t)[l].offset
 963              : t->end_offset) -
 964             rw_aux_tree(b, t)[l - 1].offset >
 965             L1_CACHE_BYTES / sizeof(u64)) {
 966                 struct bkey_packed *start = rw_aux_to_bkey(b, t, l - 1);
 967                 struct bkey_packed *end = l < t->size
 968                         ? rw_aux_to_bkey(b, t, l)
 969                         : btree_bkey_last(b, t);
 970                 struct bkey_packed *k = start;
 971
 972                 while (1) {
 973                         k = bkey_p_next(k);
 974                         if (k == end)
 975                                 break;
 976
 977                         if ((void *) k - (void *) start >= L1_CACHE_BYTES) {
 978                                 memmove(&rw_aux_tree(b, t)[l + 1],
 979                                         &rw_aux_tree(b, t)[l],
 980                                         (void *) &rw_aux_tree(b, t)[t->size] -
 981                                         (void *) &rw_aux_tree(b, t)[l]);
 982                                 t->size++;
 983                                 rw_aux_tree_set(b, t, l, k);
 984                                 break;
 985                         }
 986                 }
 987         }
 988
 989         bch2_bset_verify_rw_aux_tree(b, t);
 990         bset_aux_tree_verify(b);
 991 }
 992
 993 void bch2_bset_insert(struct btree *b,
 994                       struct btree_node_iter *iter,
 995                       struct bkey_packed *where,
 996                       struct bkey_i *insert,
 997                       unsigned clobber_u64s)
 998 {
 999         struct bkey_format *f = &b->format;
1000         struct bset_tree *t = bset_tree_last(b);
1001         struct bkey_packed packed, *src = bkey_to_packed(insert);
1002
1003         bch2_bset_verify_rw_aux_tree(b, t);
1004         bch2_verify_insert_pos(b, where, bkey_to_packed(insert), clobber_u64s);
1005
1006         if (bch2_bkey_pack_key(&packed, &insert->k, f))
1007                 src = &packed;
1008
1009         if (!bkey_deleted(&insert->k))
1010                 btree_keys_account_key_add(&b->nr, t - b->set, src);
1011
1012         if (src->u64s != clobber_u64s) {
1013                 u64 *src_p = where->_data + clobber_u64s;
1014                 u64 *dst_p = where->_data + src->u64s;
1015
1016                 EBUG_ON((int) le16_to_cpu(bset(b, t)->u64s) <
1017                         (int) clobber_u64s - src->u64s);
1018
1019                 memmove_u64s(dst_p, src_p, btree_bkey_last(b, t)->_data - src_p);
1020                 le16_add_cpu(&bset(b, t)->u64s, src->u64s - clobber_u64s);
1021                 set_btree_bset_end(b, t);
1022         }
1023
1024         memcpy_u64s_small(where, src,
1025                     bkeyp_key_u64s(f, src));
1026         memcpy_u64s(bkeyp_val(f, where), &insert->v,
1027                     bkeyp_val_u64s(f, src));
1028
1029         if (src->u64s != clobber_u64s)
1030                 bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, src->u64s);
1031
1032         bch2_verify_btree_nr_keys(b);
1033 }
1034
1035 void bch2_bset_delete(struct btree *b,
1036                       struct bkey_packed *where,
1037                       unsigned clobber_u64s)
1038 {
1039         struct bset_tree *t = bset_tree_last(b);
1040         u64 *src_p = where->_data + clobber_u64s;
1041         u64 *dst_p = where->_data;
1042
1043         bch2_bset_verify_rw_aux_tree(b, t);
1044
1045         EBUG_ON(le16_to_cpu(bset(b, t)->u64s) < clobber_u64s);
1046
1047         memmove_u64s_down(dst_p, src_p, btree_bkey_last(b, t)->_data - src_p);
1048         le16_add_cpu(&bset(b, t)->u64s, -clobber_u64s);
1049         set_btree_bset_end(b, t);
1050
1051         bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, 0);
1052 }
1053
1054 /* Lookup */
1055
1056 __flatten
1057 static struct bkey_packed *bset_search_write_set(const struct btree *b,
1058                                 struct bset_tree *t,
1059                                 struct bpos *search)
1060 {
1061         unsigned l = 0, r = t->size;
1062
1063         while (l + 1 != r) {
1064                 unsigned m = (l + r) >> 1;
1065
1066                 if (bpos_lt(rw_aux_tree(b, t)[m].k, *search))
1067                         l = m;
1068                 else
1069                         r = m;
1070         }
1071
1072         return rw_aux_to_bkey(b, t, l);
1073 }
1074
1075 static inline void prefetch_four_cachelines(void *p)
1076 {
1077 #ifdef CONFIG_X86_64
1078         asm("prefetcht0 (-127 + 64 * 0)(%0);"
1079             "prefetcht0 (-127 + 64 * 1)(%0);"
1080             "prefetcht0 (-127 + 64 * 2)(%0);"
1081             "prefetcht0 (-127 + 64 * 3)(%0);"
1082             :
1083             : "r" (p + 127));
1084 #else
1085         prefetch(p + L1_CACHE_BYTES * 0);
1086         prefetch(p + L1_CACHE_BYTES * 1);
1087         prefetch(p + L1_CACHE_BYTES * 2);
1088         prefetch(p + L1_CACHE_BYTES * 3);
1089 #endif
1090 }
1091
1092 static inline bool bkey_mantissa_bits_dropped(const struct btree *b,
1093                                               const struct bkey_float *f,
1094                                               unsigned idx)
1095 {
1096 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
1097         unsigned key_bits_start = b->format.key_u64s * 64 - b->nr_key_bits;
1098
1099         return f->exponent > key_bits_start;
1100 #else
1101         unsigned key_bits_end = high_bit_offset + b->nr_key_bits;
1102
1103         return f->exponent + BKEY_MANTISSA_BITS < key_bits_end;
1104 #endif
1105 }
1106
1107 __flatten
1108 static struct bkey_packed *bset_search_tree(const struct btree *b,
1109                                 const struct bset_tree *t,
1110                                 const struct bpos *search,
1111                                 const struct bkey_packed *packed_search)
1112 {
1113         struct ro_aux_tree *base = ro_aux_tree_base(b, t);
1114         struct bkey_float *f;
1115         struct bkey_packed *k;
1116         unsigned inorder, n = 1, l, r;
1117         int cmp;
1118
1119         do {
1120                 if (likely(n << 4 < t->size))
1121                         prefetch(&base->f[n << 4]);
1122
1123                 f = &base->f[n];
1124                 if (unlikely(f->exponent >= BFLOAT_FAILED))
1125                         goto slowpath;
1126
1127                 l = f->mantissa;
1128                 r = bkey_mantissa(packed_search, f, n);
1129
1130                 if (unlikely(l == r) && bkey_mantissa_bits_dropped(b, f, n))
1131                         goto slowpath;
1132
1133                 n = n * 2 + (l < r);
1134                 continue;
1135 slowpath:
1136                 k = tree_to_bkey(b, t, n);
1137                 cmp = bkey_cmp_p_or_unp(b, k, packed_search, search);
1138                 if (!cmp)
1139                         return k;
1140
1141                 n = n * 2 + (cmp < 0);
1142         } while (n < t->size);
1143
1144         inorder = __eytzinger1_to_inorder(n >> 1, t->size - 1, t->extra);
1145
1146         /*
1147          * n would have been the node we recursed to - the low bit tells us if
1148          * we recursed left or recursed right.
1149          */
1150         if (likely(!(n & 1))) {
1151                 --inorder;
1152                 if (unlikely(!inorder))
1153                         return btree_bkey_first(b, t);
1154
1155                 f = &base->f[eytzinger1_prev(n >> 1, t->size - 1)];
1156         }
1157
1158         return cacheline_to_bkey(b, t, inorder, f->key_offset);
1159 }
1160
1161 static __always_inline __flatten
1162 struct bkey_packed *__bch2_bset_search(struct btree *b,
1163                                 struct bset_tree *t,
1164                                 struct bpos *search,
1165                                 const struct bkey_packed *lossy_packed_search)
1166 {
1167
1168         /*
1169          * First, we search for a cacheline, then lastly we do a linear search
1170          * within that cacheline.
1171          *
1172          * To search for the cacheline, there's three different possibilities:
1173          *  * The set is too small to have a search tree, so we just do a linear
1174          *    search over the whole set.
1175          *  * The set is the one we're currently inserting into; keeping a full
1176          *    auxiliary search tree up to date would be too expensive, so we
1177          *    use a much simpler lookup table to do a binary search -
1178          *    bset_search_write_set().
1179          *  * Or we use the auxiliary search tree we constructed earlier -
1180          *    bset_search_tree()
1181          */
1182
1183         switch (bset_aux_tree_type(t)) {
1184         case BSET_NO_AUX_TREE:
1185                 return btree_bkey_first(b, t);
1186         case BSET_RW_AUX_TREE:
1187                 return bset_search_write_set(b, t, search);
1188         case BSET_RO_AUX_TREE:
1189                 return bset_search_tree(b, t, search, lossy_packed_search);
1190         default:
1191                 unreachable();
1192         }
1193 }
1194
1195 static __always_inline __flatten
1196 struct bkey_packed *bch2_bset_search_linear(struct btree *b,
1197                                 struct bset_tree *t,
1198                                 struct bpos *search,
1199                                 struct bkey_packed *packed_search,
1200                                 const struct bkey_packed *lossy_packed_search,
1201                                 struct bkey_packed *m)
1202 {
1203         if (lossy_packed_search)
1204                 while (m != btree_bkey_last(b, t) &&
1205                        bkey_iter_cmp_p_or_unp(b, m,
1206                                         lossy_packed_search, search) < 0)
1207                         m = bkey_p_next(m);
1208
1209         if (!packed_search)
1210                 while (m != btree_bkey_last(b, t) &&
1211                        bkey_iter_pos_cmp(b, m, search) < 0)
1212                         m = bkey_p_next(m);
1213
1214         if (bch2_expensive_debug_checks) {
1215                 struct bkey_packed *prev = bch2_bkey_prev_all(b, t, m);
1216
1217                 BUG_ON(prev &&
1218                        bkey_iter_cmp_p_or_unp(b, prev,
1219                                         packed_search, search) >= 0);
1220         }
1221
1222         return m;
1223 }
1224
1225 /* Btree node iterator */
1226
1227 static inline void __bch2_btree_node_iter_push(struct btree_node_iter *iter,
1228                               struct btree *b,
1229                               const struct bkey_packed *k,
1230                               const struct bkey_packed *end)
1231 {
1232         if (k != end) {
1233                 struct btree_node_iter_set *pos;
1234
1235                 btree_node_iter_for_each(iter, pos)
1236                         ;
1237
1238                 BUG_ON(pos >= iter->data + ARRAY_SIZE(iter->data));
1239                 *pos = (struct btree_node_iter_set) {
1240                         __btree_node_key_to_offset(b, k),
1241                         __btree_node_key_to_offset(b, end)
1242                 };
1243         }
1244 }
1245
1246 void bch2_btree_node_iter_push(struct btree_node_iter *iter,
1247                                struct btree *b,
1248                                const struct bkey_packed *k,
1249                                const struct bkey_packed *end)
1250 {
1251         __bch2_btree_node_iter_push(iter, b, k, end);
1252         bch2_btree_node_iter_sort(iter, b);
1253 }
1254
1255 noinline __flatten __cold
1256 static void btree_node_iter_init_pack_failed(struct btree_node_iter *iter,
1257                               struct btree *b, struct bpos *search)
1258 {
1259         struct bkey_packed *k;
1260
1261         trace_bkey_pack_pos_fail(search);
1262
1263         bch2_btree_node_iter_init_from_start(iter, b);
1264
1265         while ((k = bch2_btree_node_iter_peek(iter, b)) &&
1266                bkey_iter_pos_cmp(b, k, search) < 0)
1267                 bch2_btree_node_iter_advance(iter, b);
1268 }
1269
1270 /**
1271  * bch_btree_node_iter_init - initialize a btree node iterator, starting from a
1272  * given position
1273  *
1274  * Main entry point to the lookup code for individual btree nodes:
1275  *
1276  * NOTE:
1277  *
1278  * When you don't filter out deleted keys, btree nodes _do_ contain duplicate
1279  * keys. This doesn't matter for most code, but it does matter for lookups.
1280  *
1281  * Some adjacent keys with a string of equal keys:
1282  *      i j k k k k l m
1283  *
1284  * If you search for k, the lookup code isn't guaranteed to return you any
1285  * specific k. The lookup code is conceptually doing a binary search and
1286  * iterating backwards is very expensive so if the pivot happens to land at the
1287  * last k that's what you'll get.
1288  *
1289  * This works out ok, but it's something to be aware of:
1290  *
1291  *  - For non extents, we guarantee that the live key comes last - see
1292  *    btree_node_iter_cmp(), keys_out_of_order(). So the duplicates you don't
1293  *    see will only be deleted keys you don't care about.
1294  *
1295  *  - For extents, deleted keys sort last (see the comment at the top of this
1296  *    file). But when you're searching for extents, you actually want the first
1297  *    key strictly greater than your search key - an extent that compares equal
1298  *    to the search key is going to have 0 sectors after the search key.
1299  *
1300  *    But this does mean that we can't just search for
1301  *    bpos_successor(start_of_range) to get the first extent that overlaps with
1302  *    the range we want - if we're unlucky and there's an extent that ends
1303  *    exactly where we searched, then there could be a deleted key at the same
1304  *    position and we'd get that when we search instead of the preceding extent
1305  *    we needed.
1306  *
1307  *    So we've got to search for start_of_range, then after the lookup iterate
1308  *    past any extents that compare equal to the position we searched for.
1309  */
1310 __flatten
1311 void bch2_btree_node_iter_init(struct btree_node_iter *iter,
1312                                struct btree *b, struct bpos *search)
1313 {
1314         struct bkey_packed p, *packed_search = NULL;
1315         struct btree_node_iter_set *pos = iter->data;
1316         struct bkey_packed *k[MAX_BSETS];
1317         unsigned i;
1318
1319         EBUG_ON(bpos_lt(*search, b->data->min_key));
1320         EBUG_ON(bpos_gt(*search, b->data->max_key));
1321         bset_aux_tree_verify(b);
1322
1323         memset(iter, 0, sizeof(*iter));
1324
1325         switch (bch2_bkey_pack_pos_lossy(&p, *search, b)) {
1326         case BKEY_PACK_POS_EXACT:
1327                 packed_search = &p;
1328                 break;
1329         case BKEY_PACK_POS_SMALLER:
1330                 packed_search = NULL;
1331                 break;
1332         case BKEY_PACK_POS_FAIL:
1333                 btree_node_iter_init_pack_failed(iter, b, search);
1334                 return;
1335         }
1336
1337         for (i = 0; i < b->nsets; i++) {
1338                 k[i] = __bch2_bset_search(b, b->set + i, search, &p);
1339                 prefetch_four_cachelines(k[i]);
1340         }
1341
1342         for (i = 0; i < b->nsets; i++) {
1343                 struct bset_tree *t = b->set + i;
1344                 struct bkey_packed *end = btree_bkey_last(b, t);
1345
1346                 k[i] = bch2_bset_search_linear(b, t, search,
1347                                                packed_search, &p, k[i]);
1348                 if (k[i] != end)
1349                         *pos++ = (struct btree_node_iter_set) {
1350                                 __btree_node_key_to_offset(b, k[i]),
1351                                 __btree_node_key_to_offset(b, end)
1352                         };
1353         }
1354
1355         bch2_btree_node_iter_sort(iter, b);
1356 }
1357
1358 void bch2_btree_node_iter_init_from_start(struct btree_node_iter *iter,
1359                                           struct btree *b)
1360 {
1361         struct bset_tree *t;
1362
1363         memset(iter, 0, sizeof(*iter));
1364
1365         for_each_bset(b, t)
1366                 __bch2_btree_node_iter_push(iter, b,
1367                                            btree_bkey_first(b, t),
1368                                            btree_bkey_last(b, t));
1369         bch2_btree_node_iter_sort(iter, b);
1370 }
1371
1372 struct bkey_packed *bch2_btree_node_iter_bset_pos(struct btree_node_iter *iter,
1373                                                   struct btree *b,
1374                                                   struct bset_tree *t)
1375 {
1376         struct btree_node_iter_set *set;
1377
1378         btree_node_iter_for_each(iter, set)
1379                 if (set->end == t->end_offset)
1380                         return __btree_node_offset_to_key(b, set->k);
1381
1382         return btree_bkey_last(b, t);
1383 }
1384
1385 static inline bool btree_node_iter_sort_two(struct btree_node_iter *iter,
1386                                             struct btree *b,
1387                                             unsigned first)
1388 {
1389         bool ret;
1390
1391         if ((ret = (btree_node_iter_cmp(b,
1392                                         iter->data[first],
1393                                         iter->data[first + 1]) > 0)))
1394                 swap(iter->data[first], iter->data[first + 1]);
1395         return ret;
1396 }
1397
1398 void bch2_btree_node_iter_sort(struct btree_node_iter *iter,
1399                                struct btree *b)
1400 {
1401         /* unrolled bubble sort: */
1402
1403         if (!__btree_node_iter_set_end(iter, 2)) {
1404                 btree_node_iter_sort_two(iter, b, 0);
1405                 btree_node_iter_sort_two(iter, b, 1);
1406         }
1407
1408         if (!__btree_node_iter_set_end(iter, 1))
1409                 btree_node_iter_sort_two(iter, b, 0);
1410 }
1411
1412 void bch2_btree_node_iter_set_drop(struct btree_node_iter *iter,
1413                                    struct btree_node_iter_set *set)
1414 {
1415         struct btree_node_iter_set *last =
1416                 iter->data + ARRAY_SIZE(iter->data) - 1;
1417
1418         memmove(&set[0], &set[1], (void *) last - (void *) set);
1419         *last = (struct btree_node_iter_set) { 0, 0 };
1420 }
1421
1422 static inline void __bch2_btree_node_iter_advance(struct btree_node_iter *iter,
1423                                                   struct btree *b)
1424 {
1425         iter->data->k += __bch2_btree_node_iter_peek_all(iter, b)->u64s;
1426
1427         EBUG_ON(iter->data->k > iter->data->end);
1428
1429         if (unlikely(__btree_node_iter_set_end(iter, 0))) {
1430                 /* avoid an expensive memmove call: */
1431                 iter->data[0] = iter->data[1];
1432                 iter->data[1] = iter->data[2];
1433                 iter->data[2] = (struct btree_node_iter_set) { 0, 0 };
1434                 return;
1435         }
1436
1437         if (__btree_node_iter_set_end(iter, 1))
1438                 return;
1439
1440         if (!btree_node_iter_sort_two(iter, b, 0))
1441                 return;
1442
1443         if (__btree_node_iter_set_end(iter, 2))
1444                 return;
1445
1446         btree_node_iter_sort_two(iter, b, 1);
1447 }
1448
1449 void bch2_btree_node_iter_advance(struct btree_node_iter *iter,
1450                                   struct btree *b)
1451 {
1452         if (bch2_expensive_debug_checks) {
1453                 bch2_btree_node_iter_verify(iter, b);
1454                 bch2_btree_node_iter_next_check(iter, b);
1455         }
1456
1457         __bch2_btree_node_iter_advance(iter, b);
1458 }
1459
1460 /*
1461  * Expensive:
1462  */
1463 struct bkey_packed *bch2_btree_node_iter_prev_all(struct btree_node_iter *iter,
1464                                                   struct btree *b)
1465 {
1466         struct bkey_packed *k, *prev = NULL;
1467         struct btree_node_iter_set *set;
1468         struct bset_tree *t;
1469         unsigned end = 0;
1470
1471         if (bch2_expensive_debug_checks)
1472                 bch2_btree_node_iter_verify(iter, b);
1473
1474         for_each_bset(b, t) {
1475                 k = bch2_bkey_prev_all(b, t,
1476                         bch2_btree_node_iter_bset_pos(iter, b, t));
1477                 if (k &&
1478                     (!prev || bkey_iter_cmp(b, k, prev) > 0)) {
1479                         prev = k;
1480                         end = t->end_offset;
1481                 }
1482         }
1483
1484         if (!prev)
1485                 return NULL;
1486
1487         /*
1488          * We're manually memmoving instead of just calling sort() to ensure the
1489          * prev we picked ends up in slot 0 - sort won't necessarily put it
1490          * there because of duplicate deleted keys:
1491          */
1492         btree_node_iter_for_each(iter, set)
1493                 if (set->end == end)
1494                         goto found;
1495
1496         BUG_ON(set != &iter->data[__btree_node_iter_used(iter)]);
1497 found:
1498         BUG_ON(set >= iter->data + ARRAY_SIZE(iter->data));
1499
1500         memmove(&iter->data[1],
1501                 &iter->data[0],
1502                 (void *) set - (void *) &iter->data[0]);
1503
1504         iter->data[0].k = __btree_node_key_to_offset(b, prev);
1505         iter->data[0].end = end;
1506
1507         if (bch2_expensive_debug_checks)
1508                 bch2_btree_node_iter_verify(iter, b);
1509         return prev;
1510 }
1511
1512 struct bkey_packed *bch2_btree_node_iter_prev(struct btree_node_iter *iter,
1513                                               struct btree *b)
1514 {
1515         struct bkey_packed *prev;
1516
1517         do {
1518                 prev = bch2_btree_node_iter_prev_all(iter, b);
1519         } while (prev && bkey_deleted(prev));
1520
1521         return prev;
1522 }
1523
1524 struct bkey_s_c bch2_btree_node_iter_peek_unpack(struct btree_node_iter *iter,
1525                                                  struct btree *b,
1526                                                  struct bkey *u)
1527 {
1528         struct bkey_packed *k = bch2_btree_node_iter_peek(iter, b);
1529
1530         return k ? bkey_disassemble(b, k, u) : bkey_s_c_null;
1531 }
1532
1533 /* Mergesort */
1534
1535 void bch2_btree_keys_stats(const struct btree *b, struct bset_stats *stats)
1536 {
1537         const struct bset_tree *t;
1538
1539         for_each_bset(b, t) {
1540                 enum bset_aux_tree_type type = bset_aux_tree_type(t);
1541                 size_t j;
1542
1543                 stats->sets[type].nr++;
1544                 stats->sets[type].bytes += le16_to_cpu(bset(b, t)->u64s) *
1545                         sizeof(u64);
1546
1547                 if (bset_has_ro_aux_tree(t)) {
1548                         stats->floats += t->size - 1;
1549
1550                         for (j = 1; j < t->size; j++)
1551                                 stats->failed +=
1552                                         bkey_float(b, t, j)->exponent ==
1553                                         BFLOAT_FAILED;
1554                 }
1555         }
1556 }
1557
1558 void bch2_bfloat_to_text(struct printbuf *out, struct btree *b,
1559                          struct bkey_packed *k)
1560 {
1561         struct bset_tree *t = bch2_bkey_to_bset(b, k);
1562         struct bkey uk;
1563         unsigned j, inorder;
1564
1565         if (!bset_has_ro_aux_tree(t))
1566                 return;
1567
1568         inorder = bkey_to_cacheline(b, t, k);
1569         if (!inorder || inorder >= t->size)
1570                 return;
1571
1572         j = __inorder_to_eytzinger1(inorder, t->size - 1, t->extra);
1573         if (k != tree_to_bkey(b, t, j))
1574                 return;
1575
1576         switch (bkey_float(b, t, j)->exponent) {
1577         case BFLOAT_FAILED:
1578                 uk = bkey_unpack_key(b, k);
1579                 prt_printf(out,
1580                        "    failed unpacked at depth %u\n"
1581                        "\t",
1582                        ilog2(j));
1583                 bch2_bpos_to_text(out, uk.p);
1584                 prt_printf(out, "\n");
1585                 break;
1586         }
1587 }