git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_types.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _BCACHEFS_BTREE_TYPES_H
   3 #define _BCACHEFS_BTREE_TYPES_H
   4
   5 #include <linux/list.h>
   6 #include <linux/rhashtable.h>
   7 #include <linux/six.h>
   8
   9 #include "bkey_methods.h"
  10 #include "buckets_types.h"
  11 #include "journal_types.h"
  12
  13 struct open_bucket;
  14 struct btree_update;
  15 struct btree_trans;
  16
  17 #define MAX_BSETS               3U
  18
  19 struct btree_nr_keys {
  20
  21         /*
  22          * Amount of live metadata (i.e. size of node after a compaction) in
  23          * units of u64s
  24          */
  25         u16                     live_u64s;
  26         u16                     bset_u64s[MAX_BSETS];
  27
  28         /* live keys only: */
  29         u16                     packed_keys;
  30         u16                     unpacked_keys;
  31 };
  32
  33 struct bset_tree {
  34         /*
  35          * We construct a binary tree in an array as if the array
  36          * started at 1, so that things line up on the same cachelines
  37          * better: see comments in bset.c at cacheline_to_bkey() for
  38          * details
  39          */
  40
  41         /* size of the binary tree and prev array */
  42         u16                     size;
  43
  44         /* function of size - precalculated for to_inorder() */
  45         u16                     extra;
  46
  47         u16                     data_offset;
  48         u16                     aux_data_offset;
  49         u16                     end_offset;
  50
  51         struct bpos             max_key;
  52 };
  53
  54 struct btree_write {
  55         struct journal_entry_pin        journal;
  56 };
  57
  58 struct btree_alloc {
  59         struct open_buckets     ob;
  60         BKEY_PADDED(k);
  61 };
  62
  63 struct btree {
  64         /* Hottest entries first */
  65         struct rhash_head       hash;
  66         u64                     hash_val;
  67
  68         struct six_lock         lock;
  69
  70         unsigned long           flags;
  71         u16                     written;
  72         u8                      level;
  73         u8                      btree_id;
  74         u8                      nsets;
  75         u8                      nr_key_bits;
  76
  77         struct bkey_format      format;
  78
  79         struct btree_node       *data;
  80         void                    *aux_data;
  81
  82         /*
  83          * Sets of sorted keys - the real btree node - plus a binary search tree
  84          *
  85          * set[0] is special; set[0]->tree, set[0]->prev and set[0]->data point
  86          * to the memory we have allocated for this btree node. Additionally,
  87          * set[0]->data points to the entire btree node as it exists on disk.
  88          */
  89         struct bset_tree        set[MAX_BSETS];
  90
  91         struct btree_nr_keys    nr;
  92         u16                     sib_u64s[2];
  93         u16                     whiteout_u64s;
  94         u8                      page_order;
  95         u8                      unpack_fn_len;
  96
  97         /*
  98          * XXX: add a delete sequence number, so when bch2_btree_node_relock()
  99          * fails because the lock sequence number has changed - i.e. the
 100          * contents were modified - we can still relock the node if it's still
 101          * the one we want, without redoing the traversal
 102          */
 103
 104         /*
 105          * For asynchronous splits/interior node updates:
 106          * When we do a split, we allocate new child nodes and update the parent
 107          * node to point to them: we update the parent in memory immediately,
 108          * but then we must wait until the children have been written out before
 109          * the update to the parent can be written - this is a list of the
 110          * btree_updates that are blocking this node from being
 111          * written:
 112          */
 113         struct list_head        write_blocked;
 114
 115         /*
 116          * Also for asynchronous splits/interior node updates:
 117          * If a btree node isn't reachable yet, we don't want to kick off
 118          * another write - because that write also won't yet be reachable and
 119          * marking it as completed before it's reachable would be incorrect:
 120          */
 121         unsigned long           will_make_reachable;
 122
 123         struct open_buckets     ob;
 124
 125         /* lru list */
 126         struct list_head        list;
 127
 128         struct btree_write      writes[2];
 129
 130 #ifdef CONFIG_BCACHEFS_DEBUG
 131         bool                    *expensive_debug_checks;
 132 #endif
 133
 134         /* Key/pointer for this btree node */
 135         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 136 };
 137
 138 struct btree_cache {
 139         struct rhashtable       table;
 140         bool                    table_init_done;
 141         /*
 142          * We never free a struct btree, except on shutdown - we just put it on
 143          * the btree_cache_freed list and reuse it later. This simplifies the
 144          * code, and it doesn't cost us much memory as the memory usage is
 145          * dominated by buffers that hold the actual btree node data and those
 146          * can be freed - and the number of struct btrees allocated is
 147          * effectively bounded.
 148          *
 149          * btree_cache_freeable effectively is a small cache - we use it because
 150          * high order page allocations can be rather expensive, and it's quite
 151          * common to delete and allocate btree nodes in quick succession. It
 152          * should never grow past ~2-3 nodes in practice.
 153          */
 154         struct mutex            lock;
 155         struct list_head        live;
 156         struct list_head        freeable;
 157         struct list_head        freed;
 158
 159         /* Number of elements in live + freeable lists */
 160         unsigned                used;
 161         unsigned                reserve;
 162         struct shrinker         shrink;
 163
 164         /*
 165          * If we need to allocate memory for a new btree node and that
 166          * allocation fails, we can cannibalize another node in the btree cache
 167          * to satisfy the allocation - lock to guarantee only one thread does
 168          * this at a time:
 169          */
 170         struct task_struct      *alloc_lock;
 171         struct closure_waitlist alloc_wait;
 172 };
 173
 174 struct btree_node_iter {
 175         struct btree_node_iter_set {
 176                 u16     k, end;
 177         } data[MAX_BSETS];
 178 };
 179
 180 enum btree_iter_type {
 181         BTREE_ITER_KEYS,
 182         BTREE_ITER_NODES,
 183 };
 184
 185 #define BTREE_ITER_TYPE                 ((1 << 2) - 1)
 186
 187 /*
 188  * Iterate over all possible positions, synthesizing deleted keys for holes:
 189  */
 190 #define BTREE_ITER_SLOTS                (1 << 2)
 191 /*
 192  * Indicates that intent locks should be taken on leaf nodes, because we expect
 193  * to be doing updates:
 194  */
 195 #define BTREE_ITER_INTENT               (1 << 3)
 196 /*
 197  * Causes the btree iterator code to prefetch additional btree nodes from disk:
 198  */
 199 #define BTREE_ITER_PREFETCH             (1 << 4)
 200 /*
 201  * Indicates that this iterator should not be reused until transaction commit,
 202  * either because a pending update references it or because the update depends
 203  * on that particular key being locked (e.g. by the str_hash code, for hash
 204  * table consistency)
 205  */
 206 #define BTREE_ITER_KEEP_UNTIL_COMMIT    (1 << 5)
 207 /*
 208  * Used in bch2_btree_iter_traverse(), to indicate whether we're searching for
 209  * @pos or the first key strictly greater than @pos
 210  */
 211 #define BTREE_ITER_IS_EXTENTS           (1 << 6)
 212 #define BTREE_ITER_ERROR                (1 << 7)
 213 #define BTREE_ITER_SET_POS_AFTER_COMMIT (1 << 8)
 214
 215 enum btree_iter_uptodate {
 216         BTREE_ITER_UPTODATE             = 0,
 217         BTREE_ITER_NEED_PEEK            = 1,
 218         BTREE_ITER_NEED_RELOCK          = 2,
 219         BTREE_ITER_NEED_TRAVERSE        = 3,
 220 };
 221
 222 /*
 223  * @pos                 - iterator's current position
 224  * @level               - current btree depth
 225  * @locks_want          - btree level below which we start taking intent locks
 226  * @nodes_locked        - bitmask indicating which nodes in @nodes are locked
 227  * @nodes_intent_locked - bitmask indicating which locks are intent locks
 228  */
 229 struct btree_iter {
 230         struct btree_trans      *trans;
 231         struct bpos             pos;
 232         struct bpos             pos_after_commit;
 233
 234         u16                     flags;
 235         u8                      idx;
 236
 237         enum btree_id           btree_id:4;
 238         enum btree_iter_uptodate uptodate:4;
 239         unsigned                level:4,
 240                                 min_depth:4,
 241                                 locks_want:4,
 242                                 nodes_locked:4,
 243                                 nodes_intent_locked:4;
 244
 245         struct btree_iter_level {
 246                 struct btree    *b;
 247                 struct btree_node_iter iter;
 248                 u32             lock_seq;
 249         }                       l[BTREE_MAX_DEPTH];
 250
 251         /*
 252          * Current unpacked key - so that bch2_btree_iter_next()/
 253          * bch2_btree_iter_next_slot() can correctly advance pos.
 254          */
 255         struct bkey             k;
 256         unsigned long           ip_allocated;
 257 };
 258
 259 static inline enum btree_iter_type btree_iter_type(struct btree_iter *iter)
 260 {
 261         return iter->flags & BTREE_ITER_TYPE;
 262 }
 263
 264 static inline struct btree_iter_level *iter_l(struct btree_iter *iter)
 265 {
 266         return iter->l + iter->level;
 267 }
 268
 269 struct btree_insert_entry {
 270         unsigned                trigger_flags;
 271         unsigned                trans_triggers_run:1;
 272         struct bkey_i           *k;
 273         struct btree_iter       *iter;
 274 };
 275
 276 #ifndef CONFIG_LOCKDEP
 277 #define BTREE_ITER_MAX          64
 278 #else
 279 #define BTREE_ITER_MAX          32
 280 #endif
 281
 282 struct btree_trans {
 283         struct bch_fs           *c;
 284 #ifdef CONFIG_BCACHEFS_DEBUG
 285         struct list_head        list;
 286         struct btree            *locking;
 287         pid_t                   pid;
 288 #endif
 289         unsigned long           ip;
 290
 291         u64                     iters_linked;
 292         u64                     iters_live;
 293         u64                     iters_touched;
 294
 295         u8                      nr_iters;
 296         u8                      nr_updates;
 297         u8                      nr_updates2;
 298         u8                      size;
 299         unsigned                used_mempool:1;
 300         unsigned                error:1;
 301         unsigned                nounlock:1;
 302         unsigned                need_reset:1;
 303
 304         unsigned                mem_top;
 305         unsigned                mem_bytes;
 306         void                    *mem;
 307
 308         struct btree_iter       *iters;
 309         struct btree_insert_entry *updates;
 310         struct btree_insert_entry *updates2;
 311
 312         /* update path: */
 313         struct jset_entry       *extra_journal_entries;
 314         unsigned                extra_journal_entry_u64s;
 315         struct journal_entry_pin *journal_pin;
 316
 317         struct journal_res      journal_res;
 318         struct journal_preres   journal_preres;
 319         u64                     *journal_seq;
 320         struct disk_reservation *disk_res;
 321         unsigned                flags;
 322         unsigned                journal_u64s;
 323         unsigned                journal_preres_u64s;
 324         struct replicas_delta_list *fs_usage_deltas;
 325
 326         struct btree_iter       iters_onstack[2];
 327         struct btree_insert_entry updates_onstack[2];
 328         struct btree_insert_entry updates2_onstack[2];
 329 };
 330
 331 #define BTREE_FLAG(flag)                                                \
 332 static inline bool btree_node_ ## flag(struct btree *b)                 \
 333 {       return test_bit(BTREE_NODE_ ## flag, &b->flags); }              \
 334                                                                         \
 335 static inline void set_btree_node_ ## flag(struct btree *b)             \
 336 {       set_bit(BTREE_NODE_ ## flag, &b->flags); }                      \
 337                                                                         \
 338 static inline void clear_btree_node_ ## flag(struct btree *b)           \
 339 {       clear_bit(BTREE_NODE_ ## flag, &b->flags); }
 340
 341 enum btree_flags {
 342         BTREE_NODE_read_in_flight,
 343         BTREE_NODE_read_error,
 344         BTREE_NODE_dirty,
 345         BTREE_NODE_need_write,
 346         BTREE_NODE_noevict,
 347         BTREE_NODE_write_idx,
 348         BTREE_NODE_accessed,
 349         BTREE_NODE_write_in_flight,
 350         BTREE_NODE_just_written,
 351         BTREE_NODE_dying,
 352         BTREE_NODE_fake,
 353         BTREE_NODE_old_extent_overwrite,
 354 };
 355
 356 BTREE_FLAG(read_in_flight);
 357 BTREE_FLAG(read_error);
 358 BTREE_FLAG(dirty);
 359 BTREE_FLAG(need_write);
 360 BTREE_FLAG(noevict);
 361 BTREE_FLAG(write_idx);
 362 BTREE_FLAG(accessed);
 363 BTREE_FLAG(write_in_flight);
 364 BTREE_FLAG(just_written);
 365 BTREE_FLAG(dying);
 366 BTREE_FLAG(fake);
 367 BTREE_FLAG(old_extent_overwrite);
 368
 369 static inline struct btree_write *btree_current_write(struct btree *b)
 370 {
 371         return b->writes + btree_node_write_idx(b);
 372 }
 373
 374 static inline struct btree_write *btree_prev_write(struct btree *b)
 375 {
 376         return b->writes + (btree_node_write_idx(b) ^ 1);
 377 }
 378
 379 static inline struct bset_tree *bset_tree_last(struct btree *b)
 380 {
 381         EBUG_ON(!b->nsets);
 382         return b->set + b->nsets - 1;
 383 }
 384
 385 static inline void *
 386 __btree_node_offset_to_ptr(const struct btree *b, u16 offset)
 387 {
 388         return (void *) ((u64 *) b->data + 1 + offset);
 389 }
 390
 391 static inline u16
 392 __btree_node_ptr_to_offset(const struct btree *b, const void *p)
 393 {
 394         u16 ret = (u64 *) p - 1 - (u64 *) b->data;
 395
 396         EBUG_ON(__btree_node_offset_to_ptr(b, ret) != p);
 397         return ret;
 398 }
 399
 400 static inline struct bset *bset(const struct btree *b,
 401                                 const struct bset_tree *t)
 402 {
 403         return __btree_node_offset_to_ptr(b, t->data_offset);
 404 }
 405
 406 static inline void set_btree_bset_end(struct btree *b, struct bset_tree *t)
 407 {
 408         t->end_offset =
 409                 __btree_node_ptr_to_offset(b, vstruct_last(bset(b, t)));
 410 }
 411
 412 static inline void set_btree_bset(struct btree *b, struct bset_tree *t,
 413                                   const struct bset *i)
 414 {
 415         t->data_offset = __btree_node_ptr_to_offset(b, i);
 416         set_btree_bset_end(b, t);
 417 }
 418
 419 static inline struct bset *btree_bset_first(struct btree *b)
 420 {
 421         return bset(b, b->set);
 422 }
 423
 424 static inline struct bset *btree_bset_last(struct btree *b)
 425 {
 426         return bset(b, bset_tree_last(b));
 427 }
 428
 429 static inline u16
 430 __btree_node_key_to_offset(const struct btree *b, const struct bkey_packed *k)
 431 {
 432         return __btree_node_ptr_to_offset(b, k);
 433 }
 434
 435 static inline struct bkey_packed *
 436 __btree_node_offset_to_key(const struct btree *b, u16 k)
 437 {
 438         return __btree_node_offset_to_ptr(b, k);
 439 }
 440
 441 static inline unsigned btree_bkey_first_offset(const struct bset_tree *t)
 442 {
 443         return t->data_offset + offsetof(struct bset, _data) / sizeof(u64);
 444 }
 445
 446 #define btree_bkey_first(_b, _t)                                        \
 447 ({                                                                      \
 448         EBUG_ON(bset(_b, _t)->start !=                                  \
 449                 __btree_node_offset_to_key(_b, btree_bkey_first_offset(_t)));\
 450                                                                         \
 451         bset(_b, _t)->start;                                            \
 452 })
 453
 454 #define btree_bkey_last(_b, _t)                                         \
 455 ({                                                                      \
 456         EBUG_ON(__btree_node_offset_to_key(_b, (_t)->end_offset) !=     \
 457                 vstruct_last(bset(_b, _t)));                            \
 458                                                                         \
 459         __btree_node_offset_to_key(_b, (_t)->end_offset);               \
 460 })
 461
 462 static inline unsigned bset_u64s(struct bset_tree *t)
 463 {
 464         return t->end_offset - t->data_offset -
 465                 sizeof(struct bset) / sizeof(u64);
 466 }
 467
 468 static inline unsigned bset_dead_u64s(struct btree *b, struct bset_tree *t)
 469 {
 470         return bset_u64s(t) - b->nr.bset_u64s[t - b->set];
 471 }
 472
 473 static inline unsigned bset_byte_offset(struct btree *b, void *i)
 474 {
 475         return i - (void *) b->data;
 476 }
 477
 478 enum btree_node_type {
 479 #define x(kwd, val, name) BKEY_TYPE_##kwd = val,
 480         BCH_BTREE_IDS()
 481 #undef x
 482         BKEY_TYPE_BTREE,
 483 };
 484
 485 /* Type of a key in btree @id at level @level: */
 486 static inline enum btree_node_type __btree_node_type(unsigned level, enum btree_id id)
 487 {
 488         return level ? BKEY_TYPE_BTREE : (enum btree_node_type) id;
 489 }
 490
 491 /* Type of keys @b contains: */
 492 static inline enum btree_node_type btree_node_type(struct btree *b)
 493 {
 494         return __btree_node_type(b->level, b->btree_id);
 495 }
 496
 497 static inline bool btree_node_type_is_extents(enum btree_node_type type)
 498 {
 499         switch (type) {
 500         case BKEY_TYPE_EXTENTS:
 501         case BKEY_TYPE_REFLINK:
 502                 return true;
 503         default:
 504                 return false;
 505         }
 506 }
 507
 508 static inline bool btree_node_is_extents(struct btree *b)
 509 {
 510         return btree_node_type_is_extents(btree_node_type(b));
 511 }
 512
 513 #define BTREE_NODE_TYPE_HAS_TRIGGERS                    \
 514         ((1U << BKEY_TYPE_EXTENTS)|                     \
 515          (1U << BKEY_TYPE_ALLOC)|                       \
 516          (1U << BKEY_TYPE_INODES)|                      \
 517          (1U << BKEY_TYPE_REFLINK)|                     \
 518          (1U << BKEY_TYPE_EC)|                          \
 519          (1U << BKEY_TYPE_BTREE))
 520
 521 #define BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS              \
 522         ((1U << BKEY_TYPE_EXTENTS)|                     \
 523          (1U << BKEY_TYPE_INODES)|                      \
 524          (1U << BKEY_TYPE_REFLINK))
 525
 526 enum btree_trigger_flags {
 527         __BTREE_TRIGGER_NORUN,          /* Don't run triggers at all */
 528         __BTREE_TRIGGER_NOOVERWRITES,   /* Don't run triggers on overwrites */
 529
 530         __BTREE_TRIGGER_INSERT,
 531         __BTREE_TRIGGER_OVERWRITE,
 532         __BTREE_TRIGGER_OVERWRITE_SPLIT,
 533
 534         __BTREE_TRIGGER_GC,
 535         __BTREE_TRIGGER_BUCKET_INVALIDATE,
 536         __BTREE_TRIGGER_ALLOC_READ,
 537         __BTREE_TRIGGER_NOATOMIC,
 538 };
 539
 540 #define BTREE_TRIGGER_NORUN             (1U << __BTREE_TRIGGER_NORUN)
 541 #define BTREE_TRIGGER_NOOVERWRITES      (1U << __BTREE_TRIGGER_NOOVERWRITES)
 542
 543 #define BTREE_TRIGGER_INSERT            (1U << __BTREE_TRIGGER_INSERT)
 544 #define BTREE_TRIGGER_OVERWRITE         (1U << __BTREE_TRIGGER_OVERWRITE)
 545 #define BTREE_TRIGGER_OVERWRITE_SPLIT   (1U << __BTREE_TRIGGER_OVERWRITE_SPLIT)
 546
 547 #define BTREE_TRIGGER_GC                (1U << __BTREE_TRIGGER_GC)
 548 #define BTREE_TRIGGER_BUCKET_INVALIDATE (1U << __BTREE_TRIGGER_BUCKET_INVALIDATE)
 549 #define BTREE_TRIGGER_ALLOC_READ        (1U << __BTREE_TRIGGER_ALLOC_READ)
 550 #define BTREE_TRIGGER_NOATOMIC          (1U << __BTREE_TRIGGER_NOATOMIC)
 551
 552 static inline bool btree_node_type_needs_gc(enum btree_node_type type)
 553 {
 554         return BTREE_NODE_TYPE_HAS_TRIGGERS & (1U << type);
 555 }
 556
 557 struct btree_root {
 558         struct btree            *b;
 559
 560         /* On disk root - see async splits: */
 561         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 562         u8                      level;
 563         u8                      alive;
 564         s8                      error;
 565 };
 566
 567 /*
 568  * Optional hook that will be called just prior to a btree node update, when
 569  * we're holding the write lock and we know what key is about to be overwritten:
 570  */
 571
 572 enum btree_insert_ret {
 573         BTREE_INSERT_OK,
 574         /* leaf node needs to be split */
 575         BTREE_INSERT_BTREE_NODE_FULL,
 576         BTREE_INSERT_ENOSPC,
 577         BTREE_INSERT_NEED_MARK_REPLICAS,
 578         BTREE_INSERT_NEED_JOURNAL_RES,
 579 };
 580
 581 enum btree_gc_coalesce_fail_reason {
 582         BTREE_GC_COALESCE_FAIL_RESERVE_GET,
 583         BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC,
 584         BTREE_GC_COALESCE_FAIL_FORMAT_FITS,
 585 };
 586
 587 enum btree_node_sibling {
 588         btree_prev_sib,
 589         btree_next_sib,
 590 };
 591
 592 typedef struct btree_nr_keys (*sort_fix_overlapping_fn)(struct bset *,
 593                                                         struct btree *,
 594                                                         struct btree_node_iter *);
 595
 596 #endif /* _BCACHEFS_BTREE_TYPES_H */