drivers/md/bcache/bset.c

   1 /*
   2  * Code for working with individual keys, and sorted sets of keys with in a
   3  * btree node
   4  *
   5  * Copyright 2012 Google, Inc.
   6  */
   7
   8 #include "bcache.h"
   9 #include "btree.h"
  10 #include "debug.h"
  11
  12 #include <linux/random.h>
  13 #include <linux/prefetch.h>
  14
  15 /* Keylists */
  16
  17 int __bch_keylist_realloc(struct keylist *l, unsigned u64s)
  18 {
  19         size_t oldsize = bch_keylist_nkeys(l);
  20         size_t newsize = oldsize + u64s;
  21         uint64_t *old_keys = l->keys_p == l->inline_keys ? NULL : l->keys_p;
  22         uint64_t *new_keys;
  23
  24         newsize = roundup_pow_of_two(newsize);
  25
  26         if (newsize <= KEYLIST_INLINE ||
  27             roundup_pow_of_two(oldsize) == newsize)
  28                 return 0;
  29
  30         new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO);
  31
  32         if (!new_keys)
  33                 return -ENOMEM;
  34
  35         if (!old_keys)
  36                 memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize);
  37
  38         l->keys_p = new_keys;
  39         l->top_p = new_keys + oldsize;
  40
  41         return 0;
  42 }
  43
  44 struct bkey *bch_keylist_pop(struct keylist *l)
  45 {
  46         struct bkey *k = l->keys;
  47
  48         if (k == l->top)
  49                 return NULL;
  50
  51         while (bkey_next(k) != l->top)
  52                 k = bkey_next(k);
  53
  54         return l->top = k;
  55 }
  56
  57 void bch_keylist_pop_front(struct keylist *l)
  58 {
  59         l->top_p -= bkey_u64s(l->keys);
  60
  61         memmove(l->keys,
  62                 bkey_next(l->keys),
  63                 bch_keylist_bytes(l));
  64 }
  65
  66 /* Key/pointer manipulation */
  67
  68 void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
  69                               unsigned i)
  70 {
  71         BUG_ON(i > KEY_PTRS(src));
  72
  73         /* Only copy the header, key, and one pointer. */
  74         memcpy(dest, src, 2 * sizeof(uint64_t));
  75         dest->ptr[0] = src->ptr[i];
  76         SET_KEY_PTRS(dest, 1);
  77         /* We didn't copy the checksum so clear that bit. */
  78         SET_KEY_CSUM(dest, 0);
  79 }
  80
  81 bool __bch_cut_front(const struct bkey *where, struct bkey *k)
  82 {
  83         unsigned i, len = 0;
  84
  85         if (bkey_cmp(where, &START_KEY(k)) <= 0)
  86                 return false;
  87
  88         if (bkey_cmp(where, k) < 0)
  89                 len = KEY_OFFSET(k) - KEY_OFFSET(where);
  90         else
  91                 bkey_copy_key(k, where);
  92
  93         for (i = 0; i < KEY_PTRS(k); i++)
  94                 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len);
  95
  96         BUG_ON(len > KEY_SIZE(k));
  97         SET_KEY_SIZE(k, len);
  98         return true;
  99 }
 100
 101 bool __bch_cut_back(const struct bkey *where, struct bkey *k)
 102 {
 103         unsigned len = 0;
 104
 105         if (bkey_cmp(where, k) >= 0)
 106                 return false;
 107
 108         BUG_ON(KEY_INODE(where) != KEY_INODE(k));
 109
 110         if (bkey_cmp(where, &START_KEY(k)) > 0)
 111                 len = KEY_OFFSET(where) - KEY_START(k);
 112
 113         bkey_copy_key(k, where);
 114
 115         BUG_ON(len > KEY_SIZE(k));
 116         SET_KEY_SIZE(k, len);
 117         return true;
 118 }
 119
 120 /* Auxiliary search trees */
 121
 122 /* 32 bits total: */
 123 #define BKEY_MID_BITS           3
 124 #define BKEY_EXPONENT_BITS      7
 125 #define BKEY_MANTISSA_BITS      (32 - BKEY_MID_BITS - BKEY_EXPONENT_BITS)
 126 #define BKEY_MANTISSA_MASK      ((1 << BKEY_MANTISSA_BITS) - 1)
 127
 128 struct bkey_float {
 129         unsigned        exponent:BKEY_EXPONENT_BITS;
 130         unsigned        m:BKEY_MID_BITS;
 131         unsigned        mantissa:BKEY_MANTISSA_BITS;
 132 } __packed;
 133
 134 /*
 135  * BSET_CACHELINE was originally intended to match the hardware cacheline size -
 136  * it used to be 64, but I realized the lookup code would touch slightly less
 137  * memory if it was 128.
 138  *
 139  * It definites the number of bytes (in struct bset) per struct bkey_float in
 140  * the auxiliar search tree - when we're done searching the bset_float tree we
 141  * have this many bytes left that we do a linear search over.
 142  *
 143  * Since (after level 5) every level of the bset_tree is on a new cacheline,
 144  * we're touching one fewer cacheline in the bset tree in exchange for one more
 145  * cacheline in the linear search - but the linear search might stop before it
 146  * gets to the second cacheline.
 147  */
 148
 149 #define BSET_CACHELINE          128
 150
 151 /* Space required for the btree node keys */
 152 static inline size_t btree_keys_bytes(struct btree *b)
 153 {
 154         return PAGE_SIZE << b->page_order;
 155 }
 156
 157 static inline size_t btree_keys_cachelines(struct btree *b)
 158 {
 159         return btree_keys_bytes(b) / BSET_CACHELINE;
 160 }
 161
 162 /* Space required for the auxiliary search trees */
 163 static inline size_t bset_tree_bytes(struct btree *b)
 164 {
 165         return btree_keys_cachelines(b) * sizeof(struct bkey_float);
 166 }
 167
 168 /* Space required for the prev pointers */
 169 static inline size_t bset_prev_bytes(struct btree *b)
 170 {
 171         return btree_keys_cachelines(b) * sizeof(uint8_t);
 172 }
 173
 174 /* Memory allocation */
 175
 176 void bch_btree_keys_free(struct btree *b)
 177 {
 178         struct bset_tree *t = b->sets;
 179
 180         if (bset_prev_bytes(b) < PAGE_SIZE)
 181                 kfree(t->prev);
 182         else
 183                 free_pages((unsigned long) t->prev,
 184                            get_order(bset_prev_bytes(b)));
 185
 186         if (bset_tree_bytes(b) < PAGE_SIZE)
 187                 kfree(t->tree);
 188         else
 189                 free_pages((unsigned long) t->tree,
 190                            get_order(bset_tree_bytes(b)));
 191
 192         free_pages((unsigned long) t->data, b->page_order);
 193
 194         t->prev = NULL;
 195         t->tree = NULL;
 196         t->data = NULL;
 197 }
 198
 199 int bch_btree_keys_alloc(struct btree *b, unsigned page_order, gfp_t gfp)
 200 {
 201         struct bset_tree *t = b->sets;
 202
 203         BUG_ON(t->data);
 204
 205         b->page_order = page_order;
 206
 207         t->data = (void *) __get_free_pages(gfp, b->page_order);
 208         if (!t->data)
 209                 goto err;
 210
 211         t->tree = bset_tree_bytes(b) < PAGE_SIZE
 212                 ? kmalloc(bset_tree_bytes(b), gfp)
 213                 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
 214         if (!t->tree)
 215                 goto err;
 216
 217         t->prev = bset_prev_bytes(b) < PAGE_SIZE
 218                 ? kmalloc(bset_prev_bytes(b), gfp)
 219                 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
 220         if (!t->prev)
 221                 goto err;
 222
 223         return 0;
 224 err:
 225         bch_btree_keys_free(b);
 226         return -ENOMEM;
 227 }
 228
 229 /* Binary tree stuff for auxiliary search trees */
 230
 231 static unsigned inorder_next(unsigned j, unsigned size)
 232 {
 233         if (j * 2 + 1 < size) {
 234                 j = j * 2 + 1;
 235
 236                 while (j * 2 < size)
 237                         j *= 2;
 238         } else
 239                 j >>= ffz(j) + 1;
 240
 241         return j;
 242 }
 243
 244 static unsigned inorder_prev(unsigned j, unsigned size)
 245 {
 246         if (j * 2 < size) {
 247                 j = j * 2;
 248
 249                 while (j * 2 + 1 < size)
 250                         j = j * 2 + 1;
 251         } else
 252                 j >>= ffs(j);
 253
 254         return j;
 255 }
 256
 257 /* I have no idea why this code works... and I'm the one who wrote it
 258  *
 259  * However, I do know what it does:
 260  * Given a binary tree constructed in an array (i.e. how you normally implement
 261  * a heap), it converts a node in the tree - referenced by array index - to the
 262  * index it would have if you did an inorder traversal.
 263  *
 264  * Also tested for every j, size up to size somewhere around 6 million.
 265  *
 266  * The binary tree starts at array index 1, not 0
 267  * extra is a function of size:
 268  *   extra = (size - rounddown_pow_of_two(size - 1)) << 1;
 269  */
 270 static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra)
 271 {
 272         unsigned b = fls(j);
 273         unsigned shift = fls(size - 1) - b;
 274
 275         j  ^= 1U << (b - 1);
 276         j <<= 1;
 277         j  |= 1;
 278         j <<= shift;
 279
 280         if (j > extra)
 281                 j -= (j - extra) >> 1;
 282
 283         return j;
 284 }
 285
 286 static unsigned to_inorder(unsigned j, struct bset_tree *t)
 287 {
 288         return __to_inorder(j, t->size, t->extra);
 289 }
 290
 291 static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra)
 292 {
 293         unsigned shift;
 294
 295         if (j > extra)
 296                 j += j - extra;
 297
 298         shift = ffs(j);
 299
 300         j >>= shift;
 301         j  |= roundup_pow_of_two(size) >> shift;
 302
 303         return j;
 304 }
 305
 306 static unsigned inorder_to_tree(unsigned j, struct bset_tree *t)
 307 {
 308         return __inorder_to_tree(j, t->size, t->extra);
 309 }
 310
 311 #if 0
 312 void inorder_test(void)
 313 {
 314         unsigned long done = 0;
 315         ktime_t start = ktime_get();
 316
 317         for (unsigned size = 2;
 318              size < 65536000;
 319              size++) {
 320                 unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1;
 321                 unsigned i = 1, j = rounddown_pow_of_two(size - 1);
 322
 323                 if (!(size % 4096))
 324                         printk(KERN_NOTICE "loop %u, %llu per us\n", size,
 325                                done / ktime_us_delta(ktime_get(), start));
 326
 327                 while (1) {
 328                         if (__inorder_to_tree(i, size, extra) != j)
 329                                 panic("size %10u j %10u i %10u", size, j, i);
 330
 331                         if (__to_inorder(j, size, extra) != i)
 332                                 panic("size %10u j %10u i %10u", size, j, i);
 333
 334                         if (j == rounddown_pow_of_two(size) - 1)
 335                                 break;
 336
 337                         BUG_ON(inorder_prev(inorder_next(j, size), size) != j);
 338
 339                         j = inorder_next(j, size);
 340                         i++;
 341                 }
 342
 343                 done += size - 1;
 344         }
 345 }
 346 #endif
 347
 348 /*
 349  * Cacheline/offset <-> bkey pointer arithmetic:
 350  *
 351  * t->tree is a binary search tree in an array; each node corresponds to a key
 352  * in one cacheline in t->set (BSET_CACHELINE bytes).
 353  *
 354  * This means we don't have to store the full index of the key that a node in
 355  * the binary tree points to; to_inorder() gives us the cacheline, and then
 356  * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes.
 357  *
 358  * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
 359  * make this work.
 360  *
 361  * To construct the bfloat for an arbitrary key we need to know what the key
 362  * immediately preceding it is: we have to check if the two keys differ in the
 363  * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
 364  * of the previous key so we can walk backwards to it from t->tree[j]'s key.
 365  */
 366
 367 static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline,
 368                                       unsigned offset)
 369 {
 370         return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8;
 371 }
 372
 373 static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k)
 374 {
 375         return ((void *) k - (void *) t->data) / BSET_CACHELINE;
 376 }
 377
 378 static unsigned bkey_to_cacheline_offset(struct bkey *k)
 379 {
 380         return ((size_t) k & (BSET_CACHELINE - 1)) / sizeof(uint64_t);
 381 }
 382
 383 static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j)
 384 {
 385         return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m);
 386 }
 387
 388 static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j)
 389 {
 390         return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]);
 391 }
 392
 393 /*
 394  * For the write set - the one we're currently inserting keys into - we don't
 395  * maintain a full search tree, we just keep a simple lookup table in t->prev.
 396  */
 397 static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline)
 398 {
 399         return cacheline_to_bkey(t, cacheline, t->prev[cacheline]);
 400 }
 401
 402 static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift)
 403 {
 404         low >>= shift;
 405         low  |= (high << 1) << (63U - shift);
 406         return low;
 407 }
 408
 409 static inline unsigned bfloat_mantissa(const struct bkey *k,
 410                                        struct bkey_float *f)
 411 {
 412         const uint64_t *p = &k->low - (f->exponent >> 6);
 413         return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK;
 414 }
 415
 416 static void make_bfloat(struct bset_tree *t, unsigned j)
 417 {
 418         struct bkey_float *f = &t->tree[j];
 419         struct bkey *m = tree_to_bkey(t, j);
 420         struct bkey *p = tree_to_prev_bkey(t, j);
 421
 422         struct bkey *l = is_power_of_2(j)
 423                 ? t->data->start
 424                 : tree_to_prev_bkey(t, j >> ffs(j));
 425
 426         struct bkey *r = is_power_of_2(j + 1)
 427                 ? bset_bkey_idx(t->data, t->data->keys - bkey_u64s(&t->end))
 428                 : tree_to_bkey(t, j >> (ffz(j) + 1));
 429
 430         BUG_ON(m < l || m > r);
 431         BUG_ON(bkey_next(p) != m);
 432
 433         if (KEY_INODE(l) != KEY_INODE(r))
 434                 f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64;
 435         else
 436                 f->exponent = fls64(r->low ^ l->low);
 437
 438         f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0);
 439
 440         /*
 441          * Setting f->exponent = 127 flags this node as failed, and causes the
 442          * lookup code to fall back to comparing against the original key.
 443          */
 444
 445         if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f))
 446                 f->mantissa = bfloat_mantissa(m, f) - 1;
 447         else
 448                 f->exponent = 127;
 449 }
 450
 451 static void bset_alloc_tree(struct btree *b, struct bset_tree *t)
 452 {
 453         if (t != b->sets) {
 454                 unsigned j = roundup(t[-1].size,
 455                                      64 / sizeof(struct bkey_float));
 456
 457                 t->tree = t[-1].tree + j;
 458                 t->prev = t[-1].prev + j;
 459         }
 460
 461         while (t < b->sets + MAX_BSETS)
 462                 t++->size = 0;
 463 }
 464
 465 static void bch_bset_build_unwritten_tree(struct btree *b)
 466 {
 467         struct bset_tree *t = bset_tree_last(b);
 468
 469         bset_alloc_tree(b, t);
 470
 471         if (t->tree != b->sets->tree + btree_keys_cachelines(b)) {
 472                 t->prev[0] = bkey_to_cacheline_offset(t->data->start);
 473                 t->size = 1;
 474         }
 475 }
 476
 477 void bch_bset_init_next(struct btree *b, struct bset *i, uint64_t magic)
 478 {
 479         if (i != b->sets->data) {
 480                 b->sets[++b->nsets].data = i;
 481                 i->seq = b->sets->data->seq;
 482         } else
 483                 get_random_bytes(&i->seq, sizeof(uint64_t));
 484
 485         i->magic        = magic;
 486         i->version      = 0;
 487         i->keys         = 0;
 488
 489         bch_bset_build_unwritten_tree(b);
 490 }
 491
 492 static void bset_build_written_tree(struct btree *b)
 493 {
 494         struct bset_tree *t = bset_tree_last(b);
 495         struct bkey *k = t->data->start;
 496         unsigned j, cacheline = 1;
 497
 498         bset_alloc_tree(b, t);
 499
 500         t->size = min_t(unsigned,
 501                         bkey_to_cacheline(t, bset_bkey_last(t->data)),
 502                         b->sets->tree + btree_keys_cachelines(b) - t->tree);
 503
 504         if (t->size < 2) {
 505                 t->size = 0;
 506                 return;
 507         }
 508
 509         t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
 510
 511         /* First we figure out where the first key in each cacheline is */
 512         for (j = inorder_next(0, t->size);
 513              j;
 514              j = inorder_next(j, t->size)) {
 515                 while (bkey_to_cacheline(t, k) != cacheline)
 516                         k = bkey_next(k);
 517
 518                 t->prev[j] = bkey_u64s(k);
 519                 k = bkey_next(k);
 520                 cacheline++;
 521                 t->tree[j].m = bkey_to_cacheline_offset(k);
 522         }
 523
 524         while (bkey_next(k) != bset_bkey_last(t->data))
 525                 k = bkey_next(k);
 526
 527         t->end = *k;
 528
 529         /* Then we build the tree */
 530         for (j = inorder_next(0, t->size);
 531              j;
 532              j = inorder_next(j, t->size))
 533                 make_bfloat(t, j);
 534 }
 535
 536 void bch_bset_fix_invalidated_key(struct btree *b, struct bkey *k)
 537 {
 538         struct bset_tree *t;
 539         unsigned inorder, j = 1;
 540
 541         for (t = b->sets; t <= bset_tree_last(b); t++)
 542                 if (k < bset_bkey_last(t->data))
 543                         goto found_set;
 544
 545         BUG();
 546 found_set:
 547         if (!t->size || !bset_written(b, t))
 548                 return;
 549
 550         inorder = bkey_to_cacheline(t, k);
 551
 552         if (k == t->data->start)
 553                 goto fix_left;
 554
 555         if (bkey_next(k) == bset_bkey_last(t->data)) {
 556                 t->end = *k;
 557                 goto fix_right;
 558         }
 559
 560         j = inorder_to_tree(inorder, t);
 561
 562         if (j &&
 563             j < t->size &&
 564             k == tree_to_bkey(t, j))
 565 fix_left:       do {
 566                         make_bfloat(t, j);
 567                         j = j * 2;
 568                 } while (j < t->size);
 569
 570         j = inorder_to_tree(inorder + 1, t);
 571
 572         if (j &&
 573             j < t->size &&
 574             k == tree_to_prev_bkey(t, j))
 575 fix_right:      do {
 576                         make_bfloat(t, j);
 577                         j = j * 2 + 1;
 578                 } while (j < t->size);
 579 }
 580
 581 static void bch_bset_fix_lookup_table(struct btree *b,
 582                                       struct bset_tree *t,
 583                                       struct bkey *k)
 584 {
 585         unsigned shift = bkey_u64s(k);
 586         unsigned j = bkey_to_cacheline(t, k);
 587
 588         /* We're getting called from btree_split() or btree_gc, just bail out */
 589         if (!t->size)
 590                 return;
 591
 592         /* k is the key we just inserted; we need to find the entry in the
 593          * lookup table for the first key that is strictly greater than k:
 594          * it's either k's cacheline or the next one
 595          */
 596         if (j < t->size &&
 597             table_to_bkey(t, j) <= k)
 598                 j++;
 599
 600         /* Adjust all the lookup table entries, and find a new key for any that
 601          * have gotten too big
 602          */
 603         for (; j < t->size; j++) {
 604                 t->prev[j] += shift;
 605
 606                 if (t->prev[j] > 7) {
 607                         k = table_to_bkey(t, j - 1);
 608
 609                         while (k < cacheline_to_bkey(t, j, 0))
 610                                 k = bkey_next(k);
 611
 612                         t->prev[j] = bkey_to_cacheline_offset(k);
 613                 }
 614         }
 615
 616         if (t->size == b->sets->tree + btree_keys_cachelines(b) - t->tree)
 617                 return;
 618
 619         /* Possibly add a new entry to the end of the lookup table */
 620
 621         for (k = table_to_bkey(t, t->size - 1);
 622              k != bset_bkey_last(t->data);
 623              k = bkey_next(k))
 624                 if (t->size == bkey_to_cacheline(t, k)) {
 625                         t->prev[t->size] = bkey_to_cacheline_offset(k);
 626                         t->size++;
 627                 }
 628 }
 629
 630 void bch_bset_insert(struct btree *b, struct bkey *where,
 631                      struct bkey *insert)
 632 {
 633         struct bset_tree *t = bset_tree_last(b);
 634
 635         BUG_ON(t->data != write_block(b));
 636         BUG_ON(bset_byte_offset(b, t->data) +
 637                __set_bytes(t->data, t->data->keys + bkey_u64s(insert)) >
 638                PAGE_SIZE << b->page_order);
 639
 640         memmove((uint64_t *) where + bkey_u64s(insert),
 641                 where,
 642                 (void *) bset_bkey_last(t->data) - (void *) where);
 643
 644         t->data->keys += bkey_u64s(insert);
 645         bkey_copy(where, insert);
 646         bch_bset_fix_lookup_table(b, t, where);
 647 }
 648
 649 struct bset_search_iter {
 650         struct bkey *l, *r;
 651 };
 652
 653 static struct bset_search_iter bset_search_write_set(struct btree *b,
 654                                                      struct bset_tree *t,
 655                                                      const struct bkey *search)
 656 {
 657         unsigned li = 0, ri = t->size;
 658
 659         BUG_ON(!b->nsets &&
 660                t->size < bkey_to_cacheline(t, bset_bkey_last(t->data)));
 661
 662         while (li + 1 != ri) {
 663                 unsigned m = (li + ri) >> 1;
 664
 665                 if (bkey_cmp(table_to_bkey(t, m), search) > 0)
 666                         ri = m;
 667                 else
 668                         li = m;
 669         }
 670
 671         return (struct bset_search_iter) {
 672                 table_to_bkey(t, li),
 673                 ri < t->size ? table_to_bkey(t, ri) : bset_bkey_last(t->data)
 674         };
 675 }
 676
 677 static struct bset_search_iter bset_search_tree(struct btree *b,
 678                                                 struct bset_tree *t,
 679                                                 const struct bkey *search)
 680 {
 681         struct bkey *l, *r;
 682         struct bkey_float *f;
 683         unsigned inorder, j, n = 1;
 684
 685         do {
 686                 unsigned p = n << 4;
 687                 p &= ((int) (p - t->size)) >> 31;
 688
 689                 prefetch(&t->tree[p]);
 690
 691                 j = n;
 692                 f = &t->tree[j];
 693
 694                 /*
 695                  * n = (f->mantissa > bfloat_mantissa())
 696                  *      ? j * 2
 697                  *      : j * 2 + 1;
 698                  *
 699                  * We need to subtract 1 from f->mantissa for the sign bit trick
 700                  * to work  - that's done in make_bfloat()
 701                  */
 702                 if (likely(f->exponent != 127))
 703                         n = j * 2 + (((unsigned)
 704                                       (f->mantissa -
 705                                        bfloat_mantissa(search, f))) >> 31);
 706                 else
 707                         n = (bkey_cmp(tree_to_bkey(t, j), search) > 0)
 708                                 ? j * 2
 709                                 : j * 2 + 1;
 710         } while (n < t->size);
 711
 712         inorder = to_inorder(j, t);
 713
 714         /*
 715          * n would have been the node we recursed to - the low bit tells us if
 716          * we recursed left or recursed right.
 717          */
 718         if (n & 1) {
 719                 l = cacheline_to_bkey(t, inorder, f->m);
 720
 721                 if (++inorder != t->size) {
 722                         f = &t->tree[inorder_next(j, t->size)];
 723                         r = cacheline_to_bkey(t, inorder, f->m);
 724                 } else
 725                         r = bset_bkey_last(t->data);
 726         } else {
 727                 r = cacheline_to_bkey(t, inorder, f->m);
 728
 729                 if (--inorder) {
 730                         f = &t->tree[inorder_prev(j, t->size)];
 731                         l = cacheline_to_bkey(t, inorder, f->m);
 732                 } else
 733                         l = t->data->start;
 734         }
 735
 736         return (struct bset_search_iter) {l, r};
 737 }
 738
 739 struct bkey *__bch_bset_search(struct btree *b, struct bset_tree *t,
 740                                const struct bkey *search)
 741 {
 742         struct bset_search_iter i;
 743
 744         /*
 745          * First, we search for a cacheline, then lastly we do a linear search
 746          * within that cacheline.
 747          *
 748          * To search for the cacheline, there's three different possibilities:
 749          *  * The set is too small to have a search tree, so we just do a linear
 750          *    search over the whole set.
 751          *  * The set is the one we're currently inserting into; keeping a full
 752          *    auxiliary search tree up to date would be too expensive, so we
 753          *    use a much simpler lookup table to do a binary search -
 754          *    bset_search_write_set().
 755          *  * Or we use the auxiliary search tree we constructed earlier -
 756          *    bset_search_tree()
 757          */
 758
 759         if (unlikely(!t->size)) {
 760                 i.l = t->data->start;
 761                 i.r = bset_bkey_last(t->data);
 762         } else if (bset_written(b, t)) {
 763                 /*
 764                  * Each node in the auxiliary search tree covers a certain range
 765                  * of bits, and keys above and below the set it covers might
 766                  * differ outside those bits - so we have to special case the
 767                  * start and end - handle that here:
 768                  */
 769
 770                 if (unlikely(bkey_cmp(search, &t->end) >= 0))
 771                         return bset_bkey_last(t->data);
 772
 773                 if (unlikely(bkey_cmp(search, t->data->start) < 0))
 774                         return t->data->start;
 775
 776                 i = bset_search_tree(b, t, search);
 777         } else
 778                 i = bset_search_write_set(b, t, search);
 779
 780         if (expensive_debug_checks(b->c)) {
 781                 BUG_ON(bset_written(b, t) &&
 782                        i.l != t->data->start &&
 783                        bkey_cmp(tree_to_prev_bkey(t,
 784                           inorder_to_tree(bkey_to_cacheline(t, i.l), t)),
 785                                 search) > 0);
 786
 787                 BUG_ON(i.r != bset_bkey_last(t->data) &&
 788                        bkey_cmp(i.r, search) <= 0);
 789         }
 790
 791         while (likely(i.l != i.r) &&
 792                bkey_cmp(i.l, search) <= 0)
 793                 i.l = bkey_next(i.l);
 794
 795         return i.l;
 796 }
 797
 798 /* Btree iterator */
 799
 800 typedef bool (btree_iter_cmp_fn)(struct btree_iter_set,
 801                                  struct btree_iter_set);
 802
 803 static inline bool btree_iter_cmp(struct btree_iter_set l,
 804                                   struct btree_iter_set r)
 805 {
 806         return bkey_cmp(l.k, r.k) > 0;
 807 }
 808
 809 static inline bool btree_iter_end(struct btree_iter *iter)
 810 {
 811         return !iter->used;
 812 }
 813
 814 void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
 815                          struct bkey *end)
 816 {
 817         if (k != end)
 818                 BUG_ON(!heap_add(iter,
 819                                  ((struct btree_iter_set) { k, end }),
 820                                  btree_iter_cmp));
 821 }
 822
 823 static struct bkey *__bch_btree_iter_init(struct btree *b,
 824                                           struct btree_iter *iter,
 825                                           struct bkey *search,
 826                                           struct bset_tree *start)
 827 {
 828         struct bkey *ret = NULL;
 829         iter->size = ARRAY_SIZE(iter->data);
 830         iter->used = 0;
 831
 832 #ifdef CONFIG_BCACHE_DEBUG
 833         iter->b = b;
 834 #endif
 835
 836         for (; start <= &b->sets[b->nsets]; start++) {
 837                 ret = bch_bset_search(b, start, search);
 838                 bch_btree_iter_push(iter, ret, bset_bkey_last(start->data));
 839         }
 840
 841         return ret;
 842 }
 843
 844 struct bkey *bch_btree_iter_init(struct btree *b,
 845                                  struct btree_iter *iter,
 846                                  struct bkey *search)
 847 {
 848         return __bch_btree_iter_init(b, iter, search, b->sets);
 849 }
 850
 851 static inline struct bkey *__bch_btree_iter_next(struct btree_iter *iter,
 852                                                  btree_iter_cmp_fn *cmp)
 853 {
 854         struct btree_iter_set unused;
 855         struct bkey *ret = NULL;
 856
 857         if (!btree_iter_end(iter)) {
 858                 bch_btree_iter_next_check(iter);
 859
 860                 ret = iter->data->k;
 861                 iter->data->k = bkey_next(iter->data->k);
 862
 863                 if (iter->data->k > iter->data->end) {
 864                         WARN_ONCE(1, "bset was corrupt!\n");
 865                         iter->data->k = iter->data->end;
 866                 }
 867
 868                 if (iter->data->k == iter->data->end)
 869                         heap_pop(iter, unused, cmp);
 870                 else
 871                         heap_sift(iter, 0, cmp);
 872         }
 873
 874         return ret;
 875 }
 876
 877 struct bkey *bch_btree_iter_next(struct btree_iter *iter)
 878 {
 879         return __bch_btree_iter_next(iter, btree_iter_cmp);
 880
 881 }
 882
 883 struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
 884                                         struct btree *b, ptr_filter_fn fn)
 885 {
 886         struct bkey *ret;
 887
 888         do {
 889                 ret = bch_btree_iter_next(iter);
 890         } while (ret && fn(b, ret));
 891
 892         return ret;
 893 }
 894
 895 /* Mergesort */
 896
 897 void bch_bset_sort_state_free(struct bset_sort_state *state)
 898 {
 899         if (state->pool)
 900                 mempool_destroy(state->pool);
 901 }
 902
 903 int bch_bset_sort_state_init(struct bset_sort_state *state, unsigned page_order)
 904 {
 905         spin_lock_init(&state->time.lock);
 906
 907         state->page_order = page_order;
 908         state->crit_factor = int_sqrt(1 << page_order);
 909
 910         state->pool = mempool_create_page_pool(1, page_order);
 911         if (!state->pool)
 912                 return -ENOMEM;
 913
 914         return 0;
 915 }
 916
 917 static void btree_mergesort(struct btree *b, struct bset *out,
 918                             struct btree_iter *iter,
 919                             bool fixup, bool remove_stale)
 920 {
 921         int i;
 922         struct bkey *k, *last = NULL;
 923         BKEY_PADDED(k) tmp;
 924         bool (*bad)(struct btree *, const struct bkey *) = remove_stale
 925                 ? bch_ptr_bad
 926                 : bch_ptr_invalid;
 927
 928         /* Heapify the iterator, using our comparison function */
 929         for (i = iter->used / 2 - 1; i >= 0; --i)
 930                 heap_sift(iter, i, b->ops->sort_cmp);
 931
 932         while (!btree_iter_end(iter)) {
 933                 if (b->ops->sort_fixup && fixup)
 934                         k = b->ops->sort_fixup(iter, &tmp.k);
 935                 else
 936                         k = NULL;
 937
 938                 if (!k)
 939                         k = __bch_btree_iter_next(iter, b->ops->sort_cmp);
 940
 941                 if (bad(b, k))
 942                         continue;
 943
 944                 if (!last) {
 945                         last = out->start;
 946                         bkey_copy(last, k);
 947                 } else if (!bch_bkey_try_merge(b, last, k)) {
 948                         last = bkey_next(last);
 949                         bkey_copy(last, k);
 950                 }
 951         }
 952
 953         out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0;
 954
 955         pr_debug("sorted %i keys", out->keys);
 956 }
 957
 958 static void __btree_sort(struct btree *b, struct btree_iter *iter,
 959                          unsigned start, unsigned order, bool fixup,
 960                          struct bset_sort_state *state)
 961 {
 962         uint64_t start_time;
 963         bool used_mempool = false;
 964         struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOIO,
 965                                                      order);
 966         if (!out) {
 967                 BUG_ON(order > state->page_order);
 968
 969                 out = page_address(mempool_alloc(state->pool, GFP_NOIO));
 970                 used_mempool = true;
 971                 order = ilog2(bucket_pages(b->c));
 972         }
 973
 974         start_time = local_clock();
 975
 976         btree_mergesort(b, out, iter, fixup, false);
 977         b->nsets = start;
 978
 979         if (!start && order == b->page_order) {
 980                 /*
 981                  * Our temporary buffer is the same size as the btree node's
 982                  * buffer, we can just swap buffers instead of doing a big
 983                  * memcpy()
 984                  */
 985
 986                 out->magic      = bset_magic(&b->c->sb);
 987                 out->seq        = b->sets[0].data->seq;
 988                 out->version    = b->sets[0].data->version;
 989                 swap(out, b->sets[0].data);
 990         } else {
 991                 b->sets[start].data->keys = out->keys;
 992                 memcpy(b->sets[start].data->start, out->start,
 993                        (void *) bset_bkey_last(out) - (void *) out->start);
 994         }
 995
 996         if (used_mempool)
 997                 mempool_free(virt_to_page(out), state->pool);
 998         else
 999                 free_pages((unsigned long) out, order);
1000
1001         bset_build_written_tree(b);
1002
1003         if (!start)
1004                 bch_time_stats_update(&state->time, start_time);
1005 }
1006
1007 void bch_btree_sort_partial(struct btree *b, unsigned start,
1008                             struct bset_sort_state *state)
1009 {
1010         size_t order = b->page_order, keys = 0;
1011         struct btree_iter iter;
1012         int oldsize = bch_count_data(b);
1013
1014         __bch_btree_iter_init(b, &iter, NULL, &b->sets[start]);
1015
1016         BUG_ON(!bset_written(b, bset_tree_last(b)) &&
1017                (bset_tree_last(b)->size || b->nsets));
1018
1019         if (start) {
1020                 unsigned i;
1021
1022                 for (i = start; i <= b->nsets; i++)
1023                         keys += b->sets[i].data->keys;
1024
1025                 order = roundup_pow_of_two(__set_bytes(b->sets->data,
1026                                                        keys)) / PAGE_SIZE;
1027                 if (order)
1028                         order = ilog2(order);
1029         }
1030
1031         __btree_sort(b, &iter, start, order, false, state);
1032
1033         EBUG_ON(b->written && oldsize >= 0 && bch_count_data(b) != oldsize);
1034 }
1035 EXPORT_SYMBOL(bch_btree_sort_partial);
1036
1037 void bch_btree_sort_and_fix_extents(struct btree *b, struct btree_iter *iter,
1038                                     struct bset_sort_state *state)
1039 {
1040         __btree_sort(b, iter, 0, b->page_order, true, state);
1041 }
1042
1043 void bch_btree_sort_into(struct btree *b, struct btree *new,
1044                          struct bset_sort_state *state)
1045 {
1046         uint64_t start_time = local_clock();
1047
1048         struct btree_iter iter;
1049         bch_btree_iter_init(b, &iter, NULL);
1050
1051         btree_mergesort(b, new->sets->data, &iter, false, true);
1052
1053         bch_time_stats_update(&state->time, start_time);
1054
1055         new->sets->size = 0;
1056 }
1057
1058 #define SORT_CRIT       (4096 / sizeof(uint64_t))
1059
1060 void bch_btree_sort_lazy(struct btree *b, struct bset_sort_state *state)
1061 {
1062         unsigned crit = SORT_CRIT;
1063         int i;
1064
1065         /* Don't sort if nothing to do */
1066         if (!b->nsets)
1067                 goto out;
1068
1069         for (i = b->nsets - 1; i >= 0; --i) {
1070                 crit *= state->crit_factor;
1071
1072                 if (b->sets[i].data->keys < crit) {
1073                         bch_btree_sort_partial(b, i, state);
1074                         return;
1075                 }
1076         }
1077
1078         /* Sort if we'd overflow */
1079         if (b->nsets + 1 == MAX_BSETS) {
1080                 bch_btree_sort(b, state);
1081                 return;
1082         }
1083
1084 out:
1085         bset_build_written_tree(b);
1086 }
1087
1088 /* Sysfs stuff */
1089
1090 struct bset_stats {
1091         struct btree_op op;
1092         size_t nodes;
1093         size_t sets_written, sets_unwritten;
1094         size_t bytes_written, bytes_unwritten;
1095         size_t floats, failed;
1096 };
1097
1098 static int btree_bset_stats(struct btree_op *op, struct btree *b)
1099 {
1100         struct bset_stats *stats = container_of(op, struct bset_stats, op);
1101         unsigned i;
1102
1103         stats->nodes++;
1104
1105         for (i = 0; i <= b->nsets; i++) {
1106                 struct bset_tree *t = &b->sets[i];
1107                 size_t bytes = t->data->keys * sizeof(uint64_t);
1108                 size_t j;
1109
1110                 if (bset_written(b, t)) {
1111                         stats->sets_written++;
1112                         stats->bytes_written += bytes;
1113
1114                         stats->floats += t->size - 1;
1115
1116                         for (j = 1; j < t->size; j++)
1117                                 if (t->tree[j].exponent == 127)
1118                                         stats->failed++;
1119                 } else {
1120                         stats->sets_unwritten++;
1121                         stats->bytes_unwritten += bytes;
1122                 }
1123         }
1124
1125         return MAP_CONTINUE;
1126 }
1127
1128 int bch_bset_print_stats(struct cache_set *c, char *buf)
1129 {
1130         struct bset_stats t;
1131         int ret;
1132
1133         memset(&t, 0, sizeof(struct bset_stats));
1134         bch_btree_op_init(&t.op, -1);
1135
1136         ret = bch_btree_map_nodes(&t.op, c, &ZERO_KEY, btree_bset_stats);
1137         if (ret < 0)
1138                 return ret;
1139
1140         return snprintf(buf, PAGE_SIZE,
1141                         "btree nodes:           %zu\n"
1142                         "written sets:          %zu\n"
1143                         "unwritten sets:                %zu\n"
1144                         "written key bytes:     %zu\n"
1145                         "unwritten key bytes:   %zu\n"
1146                         "floats:                        %zu\n"
1147                         "failed:                        %zu\n",
1148                         t.nodes,
1149                         t.sets_written, t.sets_unwritten,
1150                         t.bytes_written, t.bytes_unwritten,
1151                         t.floats, t.failed);
1152 }