drivers/md/bcache/extents.c

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <[email protected]>
   3  *
   4  * Uses a block device as cache for other block devices; optimized for SSDs.
   5  * All allocation is done in buckets, which should match the erase block size
   6  * of the device.
   7  *
   8  * Buckets containing cached data are kept on a heap sorted by priority;
   9  * bucket priority is increased on cache hit, and periodically all the buckets
  10  * on the heap have their priority scaled down. This currently is just used as
  11  * an LRU but in the future should allow for more intelligent heuristics.
  12  *
  13  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
  14  * counter. Garbage collection is used to remove stale pointers.
  15  *
  16  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
  17  * as keys are inserted we only sort the pages that have not yet been written.
  18  * When garbage collection is run, we resort the entire node.
  19  *
  20  * All configuration is done via sysfs; see Documentation/bcache.txt.
  21  */
  22
  23 #include "bcache.h"
  24 #include "btree.h"
  25 #include "debug.h"
  26 #include "extents.h"
  27 #include "writeback.h"
  28
  29 static void sort_key_next(struct btree_iter *iter,
  30                           struct btree_iter_set *i)
  31 {
  32         i->k = bkey_next(i->k);
  33
  34         if (i->k == i->end)
  35                 *i = iter->data[--iter->used];
  36 }
  37
  38 static bool bch_key_sort_cmp(struct btree_iter_set l,
  39                              struct btree_iter_set r)
  40 {
  41         int64_t c = bkey_cmp(l.k, r.k);
  42
  43         return c ? c > 0 : l.k < r.k;
  44 }
  45
  46 static bool __ptr_invalid(struct cache_set *c, const struct bkey *k)
  47 {
  48         unsigned i;
  49
  50         for (i = 0; i < KEY_PTRS(k); i++)
  51                 if (ptr_available(c, k, i)) {
  52                         struct cache *ca = PTR_CACHE(c, k, i);
  53                         size_t bucket = PTR_BUCKET_NR(c, k, i);
  54                         size_t r = bucket_remainder(c, PTR_OFFSET(k, i));
  55
  56                         if (KEY_SIZE(k) + r > c->sb.bucket_size ||
  57                             bucket <  ca->sb.first_bucket ||
  58                             bucket >= ca->sb.nbuckets)
  59                                 return true;
  60                 }
  61
  62         return false;
  63 }
  64
  65 /* Common among btree and extent ptrs */
  66
  67 static const char *bch_ptr_status(struct cache_set *c, const struct bkey *k)
  68 {
  69         unsigned i;
  70
  71         for (i = 0; i < KEY_PTRS(k); i++)
  72                 if (ptr_available(c, k, i)) {
  73                         struct cache *ca = PTR_CACHE(c, k, i);
  74                         size_t bucket = PTR_BUCKET_NR(c, k, i);
  75                         size_t r = bucket_remainder(c, PTR_OFFSET(k, i));
  76
  77                         if (KEY_SIZE(k) + r > c->sb.bucket_size)
  78                                 return "bad, length too big";
  79                         if (bucket <  ca->sb.first_bucket)
  80                                 return "bad, short offset";
  81                         if (bucket >= ca->sb.nbuckets)
  82                                 return "bad, offset past end of device";
  83                         if (ptr_stale(c, k, i))
  84                                 return "stale";
  85                 }
  86
  87         if (!bkey_cmp(k, &ZERO_KEY))
  88                 return "bad, null key";
  89         if (!KEY_PTRS(k))
  90                 return "bad, no pointers";
  91         if (!KEY_SIZE(k))
  92                 return "zeroed key";
  93         return "";
  94 }
  95
  96 void bch_extent_to_text(char *buf, size_t size, const struct bkey *k)
  97 {
  98         unsigned i = 0;
  99         char *out = buf, *end = buf + size;
 100
 101 #define p(...)  (out += scnprintf(out, end - out, __VA_ARGS__))
 102
 103         p("%llu:%llu len %llu -> [", KEY_INODE(k), KEY_START(k), KEY_SIZE(k));
 104
 105         for (i = 0; i < KEY_PTRS(k); i++) {
 106                 if (i)
 107                         p(", ");
 108
 109                 if (PTR_DEV(k, i) == PTR_CHECK_DEV)
 110                         p("check dev");
 111                 else
 112                         p("%llu:%llu gen %llu", PTR_DEV(k, i),
 113                           PTR_OFFSET(k, i), PTR_GEN(k, i));
 114         }
 115
 116         p("]");
 117
 118         if (KEY_DIRTY(k))
 119                 p(" dirty");
 120         if (KEY_CSUM(k))
 121                 p(" cs%llu %llx", KEY_CSUM(k), k->ptr[1]);
 122 #undef p
 123 }
 124
 125 static void bch_bkey_dump(struct btree_keys *keys, const struct bkey *k)
 126 {
 127         struct btree *b = container_of(keys, struct btree, keys);
 128         unsigned j;
 129         char buf[80];
 130
 131         bch_extent_to_text(buf, sizeof(buf), k);
 132         printk(" %s", buf);
 133
 134         for (j = 0; j < KEY_PTRS(k); j++) {
 135                 size_t n = PTR_BUCKET_NR(b->c, k, j);
 136                 printk(" bucket %zu", n);
 137
 138                 if (n >= b->c->sb.first_bucket && n < b->c->sb.nbuckets)
 139                         printk(" prio %i",
 140                                PTR_BUCKET(b->c, k, j)->prio);
 141         }
 142
 143         printk(" %s\n", bch_ptr_status(b->c, k));
 144 }
 145
 146 /* Btree ptrs */
 147
 148 bool __bch_btree_ptr_invalid(struct cache_set *c, const struct bkey *k)
 149 {
 150         char buf[80];
 151
 152         if (!KEY_PTRS(k) || !KEY_SIZE(k) || KEY_DIRTY(k))
 153                 goto bad;
 154
 155         if (__ptr_invalid(c, k))
 156                 goto bad;
 157
 158         return false;
 159 bad:
 160         bch_extent_to_text(buf, sizeof(buf), k);
 161         cache_bug(c, "spotted btree ptr %s: %s", buf, bch_ptr_status(c, k));
 162         return true;
 163 }
 164
 165 static bool bch_btree_ptr_invalid(struct btree_keys *bk, const struct bkey *k)
 166 {
 167         struct btree *b = container_of(bk, struct btree, keys);
 168         return __bch_btree_ptr_invalid(b->c, k);
 169 }
 170
 171 static bool btree_ptr_bad_expensive(struct btree *b, const struct bkey *k)
 172 {
 173         unsigned i;
 174         char buf[80];
 175         struct bucket *g;
 176
 177         if (mutex_trylock(&b->c->bucket_lock)) {
 178                 for (i = 0; i < KEY_PTRS(k); i++)
 179                         if (ptr_available(b->c, k, i)) {
 180                                 g = PTR_BUCKET(b->c, k, i);
 181
 182                                 if (KEY_DIRTY(k) ||
 183                                     g->prio != BTREE_PRIO ||
 184                                     (b->c->gc_mark_valid &&
 185                                      GC_MARK(g) != GC_MARK_METADATA))
 186                                         goto err;
 187                         }
 188
 189                 mutex_unlock(&b->c->bucket_lock);
 190         }
 191
 192         return false;
 193 err:
 194         mutex_unlock(&b->c->bucket_lock);
 195         bch_extent_to_text(buf, sizeof(buf), k);
 196         btree_bug(b,
 197 "inconsistent btree pointer %s: bucket %zi pin %i prio %i gen %i last_gc %i mark %llu",
 198                   buf, PTR_BUCKET_NR(b->c, k, i), atomic_read(&g->pin),
 199                   g->prio, g->gen, g->last_gc, GC_MARK(g));
 200         return true;
 201 }
 202
 203 static bool bch_btree_ptr_bad(struct btree_keys *bk, const struct bkey *k)
 204 {
 205         struct btree *b = container_of(bk, struct btree, keys);
 206         unsigned i;
 207
 208         if (!bkey_cmp(k, &ZERO_KEY) ||
 209             !KEY_PTRS(k) ||
 210             bch_ptr_invalid(bk, k))
 211                 return true;
 212
 213         for (i = 0; i < KEY_PTRS(k); i++)
 214                 if (!ptr_available(b->c, k, i) ||
 215                     ptr_stale(b->c, k, i))
 216                         return true;
 217
 218         if (expensive_debug_checks(b->c) &&
 219             btree_ptr_bad_expensive(b, k))
 220                 return true;
 221
 222         return false;
 223 }
 224
 225 static bool bch_btree_ptr_insert_fixup(struct btree_keys *bk,
 226                                        struct bkey *insert,
 227                                        struct btree_iter *iter,
 228                                        struct bkey *replace_key)
 229 {
 230         struct btree *b = container_of(bk, struct btree, keys);
 231
 232         if (!KEY_OFFSET(insert))
 233                 btree_current_write(b)->prio_blocked++;
 234
 235         return false;
 236 }
 237
 238 const struct btree_keys_ops bch_btree_keys_ops = {
 239         .sort_cmp       = bch_key_sort_cmp,
 240         .insert_fixup   = bch_btree_ptr_insert_fixup,
 241         .key_invalid    = bch_btree_ptr_invalid,
 242         .key_bad        = bch_btree_ptr_bad,
 243         .key_to_text    = bch_extent_to_text,
 244         .key_dump       = bch_bkey_dump,
 245 };
 246
 247 /* Extents */
 248
 249 /*
 250  * Returns true if l > r - unless l == r, in which case returns true if l is
 251  * older than r.
 252  *
 253  * Necessary for btree_sort_fixup() - if there are multiple keys that compare
 254  * equal in different sets, we have to process them newest to oldest.
 255  */
 256 static bool bch_extent_sort_cmp(struct btree_iter_set l,
 257                                 struct btree_iter_set r)
 258 {
 259         int64_t c = bkey_cmp(&START_KEY(l.k), &START_KEY(r.k));
 260
 261         return c ? c > 0 : l.k < r.k;
 262 }
 263
 264 static struct bkey *bch_extent_sort_fixup(struct btree_iter *iter,
 265                                           struct bkey *tmp)
 266 {
 267         while (iter->used > 1) {
 268                 struct btree_iter_set *top = iter->data, *i = top + 1;
 269
 270                 if (iter->used > 2 &&
 271                     bch_extent_sort_cmp(i[0], i[1]))
 272                         i++;
 273
 274                 if (bkey_cmp(top->k, &START_KEY(i->k)) <= 0)
 275                         break;
 276
 277                 if (!KEY_SIZE(i->k)) {
 278                         sort_key_next(iter, i);
 279                         heap_sift(iter, i - top, bch_extent_sort_cmp);
 280                         continue;
 281                 }
 282
 283                 if (top->k > i->k) {
 284                         if (bkey_cmp(top->k, i->k) >= 0)
 285                                 sort_key_next(iter, i);
 286                         else
 287                                 bch_cut_front(top->k, i->k);
 288
 289                         heap_sift(iter, i - top, bch_extent_sort_cmp);
 290                 } else {
 291                         /* can't happen because of comparison func */
 292                         BUG_ON(!bkey_cmp(&START_KEY(top->k), &START_KEY(i->k)));
 293
 294                         if (bkey_cmp(i->k, top->k) < 0) {
 295                                 bkey_copy(tmp, top->k);
 296
 297                                 bch_cut_back(&START_KEY(i->k), tmp);
 298                                 bch_cut_front(i->k, top->k);
 299                                 heap_sift(iter, 0, bch_extent_sort_cmp);
 300
 301                                 return tmp;
 302                         } else {
 303                                 bch_cut_back(&START_KEY(i->k), top->k);
 304                         }
 305                 }
 306         }
 307
 308         return NULL;
 309 }
 310
 311 static void bch_subtract_dirty(struct bkey *k,
 312                            struct cache_set *c,
 313                            uint64_t offset,
 314                            int sectors)
 315 {
 316         if (KEY_DIRTY(k))
 317                 bcache_dev_sectors_dirty_add(c, KEY_INODE(k),
 318                                              offset, -sectors);
 319 }
 320
 321 static bool bch_extent_insert_fixup(struct btree_keys *b,
 322                                     struct bkey *insert,
 323                                     struct btree_iter *iter,
 324                                     struct bkey *replace_key)
 325 {
 326         struct cache_set *c = container_of(b, struct btree, keys)->c;
 327
 328         uint64_t old_offset;
 329         unsigned old_size, sectors_found = 0;
 330
 331         BUG_ON(!KEY_OFFSET(insert));
 332         BUG_ON(!KEY_SIZE(insert));
 333
 334         while (1) {
 335                 struct bkey *k = bch_btree_iter_next(iter);
 336                 if (!k)
 337                         break;
 338
 339                 if (bkey_cmp(&START_KEY(k), insert) >= 0) {
 340                         if (KEY_SIZE(k))
 341                                 break;
 342                         else
 343                                 continue;
 344                 }
 345
 346                 if (bkey_cmp(k, &START_KEY(insert)) <= 0)
 347                         continue;
 348
 349                 old_offset = KEY_START(k);
 350                 old_size = KEY_SIZE(k);
 351
 352                 /*
 353                  * We might overlap with 0 size extents; we can't skip these
 354                  * because if they're in the set we're inserting to we have to
 355                  * adjust them so they don't overlap with the key we're
 356                  * inserting. But we don't want to check them for replace
 357                  * operations.
 358                  */
 359
 360                 if (replace_key && KEY_SIZE(k)) {
 361                         /*
 362                          * k might have been split since we inserted/found the
 363                          * key we're replacing
 364                          */
 365                         unsigned i;
 366                         uint64_t offset = KEY_START(k) -
 367                                 KEY_START(replace_key);
 368
 369                         /* But it must be a subset of the replace key */
 370                         if (KEY_START(k) < KEY_START(replace_key) ||
 371                             KEY_OFFSET(k) > KEY_OFFSET(replace_key))
 372                                 goto check_failed;
 373
 374                         /* We didn't find a key that we were supposed to */
 375                         if (KEY_START(k) > KEY_START(insert) + sectors_found)
 376                                 goto check_failed;
 377
 378                         if (!bch_bkey_equal_header(k, replace_key))
 379                                 goto check_failed;
 380
 381                         /* skip past gen */
 382                         offset <<= 8;
 383
 384                         BUG_ON(!KEY_PTRS(replace_key));
 385
 386                         for (i = 0; i < KEY_PTRS(replace_key); i++)
 387                                 if (k->ptr[i] != replace_key->ptr[i] + offset)
 388                                         goto check_failed;
 389
 390                         sectors_found = KEY_OFFSET(k) - KEY_START(insert);
 391                 }
 392
 393                 if (bkey_cmp(insert, k) < 0 &&
 394                     bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) {
 395                         /*
 396                          * We overlapped in the middle of an existing key: that
 397                          * means we have to split the old key. But we have to do
 398                          * slightly different things depending on whether the
 399                          * old key has been written out yet.
 400                          */
 401
 402                         struct bkey *top;
 403
 404                         bch_subtract_dirty(k, c, KEY_START(insert),
 405                                        KEY_SIZE(insert));
 406
 407                         if (bkey_written(b, k)) {
 408                                 /*
 409                                  * We insert a new key to cover the top of the
 410                                  * old key, and the old key is modified in place
 411                                  * to represent the bottom split.
 412                                  *
 413                                  * It's completely arbitrary whether the new key
 414                                  * is the top or the bottom, but it has to match
 415                                  * up with what btree_sort_fixup() does - it
 416                                  * doesn't check for this kind of overlap, it
 417                                  * depends on us inserting a new key for the top
 418                                  * here.
 419                                  */
 420                                 top = bch_bset_search(b, bset_tree_last(b),
 421                                                       insert);
 422                                 bch_bset_insert(b, top, k);
 423                         } else {
 424                                 BKEY_PADDED(key) temp;
 425                                 bkey_copy(&temp.key, k);
 426                                 bch_bset_insert(b, k, &temp.key);
 427                                 top = bkey_next(k);
 428                         }
 429
 430                         bch_cut_front(insert, top);
 431                         bch_cut_back(&START_KEY(insert), k);
 432                         bch_bset_fix_invalidated_key(b, k);
 433                         goto out;
 434                 }
 435
 436                 if (bkey_cmp(insert, k) < 0) {
 437                         bch_cut_front(insert, k);
 438                 } else {
 439                         if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0)
 440                                 old_offset = KEY_START(insert);
 441
 442                         if (bkey_written(b, k) &&
 443                             bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) {
 444                                 /*
 445                                  * Completely overwrote, so we don't have to
 446                                  * invalidate the binary search tree
 447                                  */
 448                                 bch_cut_front(k, k);
 449                         } else {
 450                                 __bch_cut_back(&START_KEY(insert), k);
 451                                 bch_bset_fix_invalidated_key(b, k);
 452                         }
 453                 }
 454
 455                 bch_subtract_dirty(k, c, old_offset, old_size - KEY_SIZE(k));
 456         }
 457
 458 check_failed:
 459         if (replace_key) {
 460                 if (!sectors_found) {
 461                         return true;
 462                 } else if (sectors_found < KEY_SIZE(insert)) {
 463                         SET_KEY_OFFSET(insert, KEY_OFFSET(insert) -
 464                                        (KEY_SIZE(insert) - sectors_found));
 465                         SET_KEY_SIZE(insert, sectors_found);
 466                 }
 467         }
 468 out:
 469         if (KEY_DIRTY(insert))
 470                 bcache_dev_sectors_dirty_add(c, KEY_INODE(insert),
 471                                              KEY_START(insert),
 472                                              KEY_SIZE(insert));
 473
 474         return false;
 475 }
 476
 477 bool __bch_extent_invalid(struct cache_set *c, const struct bkey *k)
 478 {
 479         char buf[80];
 480
 481         if (!KEY_SIZE(k))
 482                 return true;
 483
 484         if (KEY_SIZE(k) > KEY_OFFSET(k))
 485                 goto bad;
 486
 487         if (__ptr_invalid(c, k))
 488                 goto bad;
 489
 490         return false;
 491 bad:
 492         bch_extent_to_text(buf, sizeof(buf), k);
 493         cache_bug(c, "spotted extent %s: %s", buf, bch_ptr_status(c, k));
 494         return true;
 495 }
 496
 497 static bool bch_extent_invalid(struct btree_keys *bk, const struct bkey *k)
 498 {
 499         struct btree *b = container_of(bk, struct btree, keys);
 500         return __bch_extent_invalid(b->c, k);
 501 }
 502
 503 static bool bch_extent_bad_expensive(struct btree *b, const struct bkey *k,
 504                                      unsigned ptr)
 505 {
 506         struct bucket *g = PTR_BUCKET(b->c, k, ptr);
 507         char buf[80];
 508
 509         if (mutex_trylock(&b->c->bucket_lock)) {
 510                 if (b->c->gc_mark_valid &&
 511                     (!GC_MARK(g) ||
 512                      GC_MARK(g) == GC_MARK_METADATA ||
 513                      (GC_MARK(g) != GC_MARK_DIRTY && KEY_DIRTY(k))))
 514                         goto err;
 515
 516                 if (g->prio == BTREE_PRIO)
 517                         goto err;
 518
 519                 mutex_unlock(&b->c->bucket_lock);
 520         }
 521
 522         return false;
 523 err:
 524         mutex_unlock(&b->c->bucket_lock);
 525         bch_extent_to_text(buf, sizeof(buf), k);
 526         btree_bug(b,
 527 "inconsistent extent pointer %s:\nbucket %zu pin %i prio %i gen %i last_gc %i mark %llu",
 528                   buf, PTR_BUCKET_NR(b->c, k, ptr), atomic_read(&g->pin),
 529                   g->prio, g->gen, g->last_gc, GC_MARK(g));
 530         return true;
 531 }
 532
 533 static bool bch_extent_bad(struct btree_keys *bk, const struct bkey *k)
 534 {
 535         struct btree *b = container_of(bk, struct btree, keys);
 536         struct bucket *g;
 537         unsigned i, stale;
 538
 539         if (!KEY_PTRS(k) ||
 540             bch_extent_invalid(bk, k))
 541                 return true;
 542
 543         for (i = 0; i < KEY_PTRS(k); i++)
 544                 if (!ptr_available(b->c, k, i))
 545                         return true;
 546
 547         if (!expensive_debug_checks(b->c) && KEY_DIRTY(k))
 548                 return false;
 549
 550         for (i = 0; i < KEY_PTRS(k); i++) {
 551                 g = PTR_BUCKET(b->c, k, i);
 552                 stale = ptr_stale(b->c, k, i);
 553
 554                 btree_bug_on(stale > 96, b,
 555                              "key too stale: %i, need_gc %u",
 556                              stale, b->c->need_gc);
 557
 558                 btree_bug_on(stale && KEY_DIRTY(k) && KEY_SIZE(k),
 559                              b, "stale dirty pointer");
 560
 561                 if (stale)
 562                         return true;
 563
 564                 if (expensive_debug_checks(b->c) &&
 565                     bch_extent_bad_expensive(b, k, i))
 566                         return true;
 567         }
 568
 569         return false;
 570 }
 571
 572 static uint64_t merge_chksums(struct bkey *l, struct bkey *r)
 573 {
 574         return (l->ptr[KEY_PTRS(l)] + r->ptr[KEY_PTRS(r)]) &
 575                 ~((uint64_t)1 << 63);
 576 }
 577
 578 static bool bch_extent_merge(struct btree_keys *bk, struct bkey *l, struct bkey *r)
 579 {
 580         struct btree *b = container_of(bk, struct btree, keys);
 581         unsigned i;
 582
 583         if (key_merging_disabled(b->c))
 584                 return false;
 585
 586         for (i = 0; i < KEY_PTRS(l); i++)
 587                 if (l->ptr[i] + PTR(0, KEY_SIZE(l), 0) != r->ptr[i] ||
 588                     PTR_BUCKET_NR(b->c, l, i) != PTR_BUCKET_NR(b->c, r, i))
 589                         return false;
 590
 591         /* Keys with no pointers aren't restricted to one bucket and could
 592          * overflow KEY_SIZE
 593          */
 594         if (KEY_SIZE(l) + KEY_SIZE(r) > USHRT_MAX) {
 595                 SET_KEY_OFFSET(l, KEY_OFFSET(l) + USHRT_MAX - KEY_SIZE(l));
 596                 SET_KEY_SIZE(l, USHRT_MAX);
 597
 598                 bch_cut_front(l, r);
 599                 return false;
 600         }
 601
 602         if (KEY_CSUM(l)) {
 603                 if (KEY_CSUM(r))
 604                         l->ptr[KEY_PTRS(l)] = merge_chksums(l, r);
 605                 else
 606                         SET_KEY_CSUM(l, 0);
 607         }
 608
 609         SET_KEY_OFFSET(l, KEY_OFFSET(l) + KEY_SIZE(r));
 610         SET_KEY_SIZE(l, KEY_SIZE(l) + KEY_SIZE(r));
 611
 612         return true;
 613 }
 614
 615 const struct btree_keys_ops bch_extent_keys_ops = {
 616         .sort_cmp       = bch_extent_sort_cmp,
 617         .sort_fixup     = bch_extent_sort_fixup,
 618         .insert_fixup   = bch_extent_insert_fixup,
 619         .key_invalid    = bch_extent_invalid,
 620         .key_bad        = bch_extent_bad,
 621         .key_merge      = bch_extent_merge,
 622         .key_to_text    = bch_extent_to_text,
 623         .key_dump       = bch_bkey_dump,
 624         .is_extents     = true,
 625 };