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

   1 #ifndef _BCACHEFS_BTREE_TYPES_H
   2 #define _BCACHEFS_BTREE_TYPES_H
   3
   4 #include <linux/list.h>
   5 #include <linux/rhashtable.h>
   6 #include <linux/six.h>
   7
   8 #include "bkey_methods.h"
   9 #include "journal_types.h"
  10
  11 struct open_bucket;
  12 struct btree_update;
  13 struct btree_trans;
  14
  15 #define MAX_BSETS               3U
  16
  17 struct btree_nr_keys {
  18
  19         /*
  20          * Amount of live metadata (i.e. size of node after a compaction) in
  21          * units of u64s
  22          */
  23         u16                     live_u64s;
  24         u16                     bset_u64s[MAX_BSETS];
  25
  26         /* live keys only: */
  27         u16                     packed_keys;
  28         u16                     unpacked_keys;
  29 };
  30
  31 struct bset_tree {
  32         /*
  33          * We construct a binary tree in an array as if the array
  34          * started at 1, so that things line up on the same cachelines
  35          * better: see comments in bset.c at cacheline_to_bkey() for
  36          * details
  37          */
  38
  39         /* size of the binary tree and prev array */
  40         u16                     size;
  41
  42         /* function of size - precalculated for to_inorder() */
  43         u16                     extra;
  44
  45         u16                     data_offset;
  46         u16                     aux_data_offset;
  47         u16                     end_offset;
  48
  49         struct bpos             max_key;
  50 };
  51
  52 struct btree_write {
  53         struct journal_entry_pin        journal;
  54         struct closure_waitlist         wait;
  55 };
  56
  57 struct btree_alloc {
  58         struct open_buckets     ob;
  59         BKEY_PADDED(k);
  60 };
  61
  62 struct btree {
  63         /* Hottest entries first */
  64         struct rhash_head       hash;
  65
  66         /* Key/pointer for this btree node */
  67         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
  68
  69         struct six_lock         lock;
  70
  71         unsigned long           flags;
  72         u16                     written;
  73         u8                      level;
  74         u8                      btree_id;
  75         u8                      nsets;
  76         u8                      nr_key_bits;
  77
  78         struct bkey_format      format;
  79
  80         struct btree_node       *data;
  81         void                    *aux_data;
  82
  83         /*
  84          * Sets of sorted keys - the real btree node - plus a binary search tree
  85          *
  86          * set[0] is special; set[0]->tree, set[0]->prev and set[0]->data point
  87          * to the memory we have allocated for this btree node. Additionally,
  88          * set[0]->data points to the entire btree node as it exists on disk.
  89          */
  90         struct bset_tree        set[MAX_BSETS];
  91
  92         struct btree_nr_keys    nr;
  93         u16                     sib_u64s[2];
  94         u16                     whiteout_u64s;
  95         u16                     uncompacted_whiteout_u64s;
  96         u8                      page_order;
  97         u8                      unpack_fn_len;
  98
  99         /*
 100          * XXX: add a delete sequence number, so when bch2_btree_node_relock()
 101          * fails because the lock sequence number has changed - i.e. the
 102          * contents were modified - we can still relock the node if it's still
 103          * the one we want, without redoing the traversal
 104          */
 105
 106         /*
 107          * For asynchronous splits/interior node updates:
 108          * When we do a split, we allocate new child nodes and update the parent
 109          * node to point to them: we update the parent in memory immediately,
 110          * but then we must wait until the children have been written out before
 111          * the update to the parent can be written - this is a list of the
 112          * btree_updates that are blocking this node from being
 113          * written:
 114          */
 115         struct list_head        write_blocked;
 116
 117         /*
 118          * Also for asynchronous splits/interior node updates:
 119          * If a btree node isn't reachable yet, we don't want to kick off
 120          * another write - because that write also won't yet be reachable and
 121          * marking it as completed before it's reachable would be incorrect:
 122          */
 123         unsigned long           will_make_reachable;
 124
 125         struct open_buckets     ob;
 126
 127         /* lru list */
 128         struct list_head        list;
 129
 130         struct btree_write      writes[2];
 131
 132 #ifdef CONFIG_BCACHEFS_DEBUG
 133         bool                    *expensive_debug_checks;
 134 #endif
 135 };
 136
 137 struct btree_cache {
 138         struct rhashtable       table;
 139         bool                    table_init_done;
 140         /*
 141          * We never free a struct btree, except on shutdown - we just put it on
 142          * the btree_cache_freed list and reuse it later. This simplifies the
 143          * code, and it doesn't cost us much memory as the memory usage is
 144          * dominated by buffers that hold the actual btree node data and those
 145          * can be freed - and the number of struct btrees allocated is
 146          * effectively bounded.
 147          *
 148          * btree_cache_freeable effectively is a small cache - we use it because
 149          * high order page allocations can be rather expensive, and it's quite
 150          * common to delete and allocate btree nodes in quick succession. It
 151          * should never grow past ~2-3 nodes in practice.
 152          */
 153         struct mutex            lock;
 154         struct list_head        live;
 155         struct list_head        freeable;
 156         struct list_head        freed;
 157
 158         /* Number of elements in live + freeable lists */
 159         unsigned                used;
 160         unsigned                reserve;
 161         struct shrinker         shrink;
 162
 163         /*
 164          * If we need to allocate memory for a new btree node and that
 165          * allocation fails, we can cannibalize another node in the btree cache
 166          * to satisfy the allocation - lock to guarantee only one thread does
 167          * this at a time:
 168          */
 169         struct task_struct      *alloc_lock;
 170         struct closure_waitlist alloc_wait;
 171 };
 172
 173 struct btree_node_iter {
 174         struct btree_node_iter_set {
 175                 u16     k, end;
 176         } data[MAX_BSETS];
 177 };
 178
 179 enum btree_iter_type {
 180         BTREE_ITER_KEYS,
 181         BTREE_ITER_SLOTS,
 182         BTREE_ITER_NODES,
 183 };
 184
 185 #define BTREE_ITER_TYPE                 ((1 << 2) - 1)
 186
 187 #define BTREE_ITER_INTENT               (1 << 2)
 188 #define BTREE_ITER_PREFETCH             (1 << 3)
 189 /*
 190  * Used in bch2_btree_iter_traverse(), to indicate whether we're searching for
 191  * @pos or the first key strictly greater than @pos
 192  */
 193 #define BTREE_ITER_IS_EXTENTS           (1 << 4)
 194 #define BTREE_ITER_ERROR                (1 << 5)
 195 #define BTREE_ITER_NOUNLOCK             (1 << 6)
 196
 197 enum btree_iter_uptodate {
 198         BTREE_ITER_UPTODATE             = 0,
 199         BTREE_ITER_NEED_PEEK            = 1,
 200         BTREE_ITER_NEED_RELOCK          = 2,
 201         BTREE_ITER_NEED_TRAVERSE        = 3,
 202 };
 203
 204 /*
 205  * @pos                 - iterator's current position
 206  * @level               - current btree depth
 207  * @locks_want          - btree level below which we start taking intent locks
 208  * @nodes_locked        - bitmask indicating which nodes in @nodes are locked
 209  * @nodes_intent_locked - bitmask indicating which locks are intent locks
 210  */
 211 struct btree_iter {
 212         u8                      idx;
 213
 214         struct btree_trans      *trans;
 215         struct bpos             pos;
 216
 217         u8                      flags;
 218         enum btree_iter_uptodate uptodate:4;
 219         enum btree_id           btree_id:4;
 220         unsigned                level:4,
 221                                 locks_want:4,
 222                                 nodes_locked:4,
 223                                 nodes_intent_locked:4;
 224
 225         struct btree_iter_level {
 226                 struct btree    *b;
 227                 struct btree_node_iter iter;
 228                 u32             lock_seq;
 229         }                       l[BTREE_MAX_DEPTH];
 230
 231         /*
 232          * Current unpacked key - so that bch2_btree_iter_next()/
 233          * bch2_btree_iter_next_slot() can correctly advance pos.
 234          */
 235         struct bkey             k;
 236
 237         u64                     id;
 238 };
 239
 240 struct deferred_update {
 241         struct journal_preres   res;
 242         struct journal_entry_pin journal;
 243
 244         spinlock_t              lock;
 245         unsigned                dirty:1;
 246
 247         u8                      allocated_u64s;
 248         enum btree_id           btree_id;
 249
 250         /* must be last: */
 251         struct bkey_i           k;
 252 };
 253
 254 struct btree_insert_entry {
 255         struct bkey_i           *k;
 256
 257         union {
 258         struct btree_iter       *iter;
 259         struct deferred_update  *d;
 260         };
 261
 262         bool                    deferred;
 263 };
 264
 265 #define BTREE_ITER_MAX          64
 266
 267 struct btree_trans {
 268         struct bch_fs           *c;
 269         size_t                  nr_restarts;
 270         u64                     commit_start;
 271
 272         u64                     iters_linked;
 273         u64                     iters_live;
 274         u64                     iters_touched;
 275         u64                     iters_unlink_on_restart;
 276         u64                     iters_unlink_on_commit;
 277
 278         u8                      nr_iters;
 279         u8                      nr_updates;
 280         u8                      size;
 281         unsigned                used_mempool:1;
 282         unsigned                error:1;
 283
 284         unsigned                mem_top;
 285         unsigned                mem_bytes;
 286         void                    *mem;
 287
 288         struct btree_iter       *iters;
 289         struct btree_insert_entry *updates;
 290
 291         /* update path: */
 292         struct journal_res      journal_res;
 293         struct journal_preres   journal_preres;
 294         u64                     *journal_seq;
 295         struct disk_reservation *disk_res;
 296         unsigned                flags;
 297
 298         struct btree_iter       iters_onstack[2];
 299         struct btree_insert_entry updates_onstack[6];
 300 };
 301
 302 #define BTREE_FLAG(flag)                                                \
 303 static inline bool btree_node_ ## flag(struct btree *b)                 \
 304 {       return test_bit(BTREE_NODE_ ## flag, &b->flags); }              \
 305                                                                         \
 306 static inline void set_btree_node_ ## flag(struct btree *b)             \
 307 {       set_bit(BTREE_NODE_ ## flag, &b->flags); }                      \
 308                                                                         \
 309 static inline void clear_btree_node_ ## flag(struct btree *b)           \
 310 {       clear_bit(BTREE_NODE_ ## flag, &b->flags); }
 311
 312 enum btree_flags {
 313         BTREE_NODE_read_in_flight,
 314         BTREE_NODE_read_error,
 315         BTREE_NODE_dirty,
 316         BTREE_NODE_need_write,
 317         BTREE_NODE_noevict,
 318         BTREE_NODE_write_idx,
 319         BTREE_NODE_accessed,
 320         BTREE_NODE_write_in_flight,
 321         BTREE_NODE_just_written,
 322         BTREE_NODE_dying,
 323         BTREE_NODE_fake,
 324 };
 325
 326 BTREE_FLAG(read_in_flight);
 327 BTREE_FLAG(read_error);
 328 BTREE_FLAG(dirty);
 329 BTREE_FLAG(need_write);
 330 BTREE_FLAG(noevict);
 331 BTREE_FLAG(write_idx);
 332 BTREE_FLAG(accessed);
 333 BTREE_FLAG(write_in_flight);
 334 BTREE_FLAG(just_written);
 335 BTREE_FLAG(dying);
 336 BTREE_FLAG(fake);
 337
 338 static inline struct btree_write *btree_current_write(struct btree *b)
 339 {
 340         return b->writes + btree_node_write_idx(b);
 341 }
 342
 343 static inline struct btree_write *btree_prev_write(struct btree *b)
 344 {
 345         return b->writes + (btree_node_write_idx(b) ^ 1);
 346 }
 347
 348 static inline struct bset_tree *bset_tree_last(struct btree *b)
 349 {
 350         EBUG_ON(!b->nsets);
 351         return b->set + b->nsets - 1;
 352 }
 353
 354 static inline void *
 355 __btree_node_offset_to_ptr(const struct btree *b, u16 offset)
 356 {
 357         return (void *) ((u64 *) b->data + 1 + offset);
 358 }
 359
 360 static inline u16
 361 __btree_node_ptr_to_offset(const struct btree *b, const void *p)
 362 {
 363         u16 ret = (u64 *) p - 1 - (u64 *) b->data;
 364
 365         EBUG_ON(__btree_node_offset_to_ptr(b, ret) != p);
 366         return ret;
 367 }
 368
 369 static inline struct bset *bset(const struct btree *b,
 370                                 const struct bset_tree *t)
 371 {
 372         return __btree_node_offset_to_ptr(b, t->data_offset);
 373 }
 374
 375 static inline void set_btree_bset_end(struct btree *b, struct bset_tree *t)
 376 {
 377         t->end_offset =
 378                 __btree_node_ptr_to_offset(b, vstruct_last(bset(b, t)));
 379 }
 380
 381 static inline void set_btree_bset(struct btree *b, struct bset_tree *t,
 382                                   const struct bset *i)
 383 {
 384         t->data_offset = __btree_node_ptr_to_offset(b, i);
 385         set_btree_bset_end(b, t);
 386 }
 387
 388 static inline struct bset *btree_bset_first(struct btree *b)
 389 {
 390         return bset(b, b->set);
 391 }
 392
 393 static inline struct bset *btree_bset_last(struct btree *b)
 394 {
 395         return bset(b, bset_tree_last(b));
 396 }
 397
 398 static inline u16
 399 __btree_node_key_to_offset(const struct btree *b, const struct bkey_packed *k)
 400 {
 401         return __btree_node_ptr_to_offset(b, k);
 402 }
 403
 404 static inline struct bkey_packed *
 405 __btree_node_offset_to_key(const struct btree *b, u16 k)
 406 {
 407         return __btree_node_offset_to_ptr(b, k);
 408 }
 409
 410 static inline unsigned btree_bkey_first_offset(const struct bset_tree *t)
 411 {
 412         return t->data_offset + offsetof(struct bset, _data) / sizeof(u64);
 413 }
 414
 415 #define btree_bkey_first(_b, _t)                                        \
 416 ({                                                                      \
 417         EBUG_ON(bset(_b, _t)->start !=                                  \
 418                 __btree_node_offset_to_key(_b, btree_bkey_first_offset(_t)));\
 419                                                                         \
 420         bset(_b, _t)->start;                                            \
 421 })
 422
 423 #define btree_bkey_last(_b, _t)                                         \
 424 ({                                                                      \
 425         EBUG_ON(__btree_node_offset_to_key(_b, (_t)->end_offset) !=     \
 426                 vstruct_last(bset(_b, _t)));                            \
 427                                                                         \
 428         __btree_node_offset_to_key(_b, (_t)->end_offset);               \
 429 })
 430
 431 static inline unsigned bset_byte_offset(struct btree *b, void *i)
 432 {
 433         return i - (void *) b->data;
 434 }
 435
 436 enum btree_node_type {
 437 #define x(kwd, val, name) BKEY_TYPE_##kwd = val,
 438         BCH_BTREE_IDS()
 439 #undef x
 440         BKEY_TYPE_BTREE,
 441 };
 442
 443 /* Type of a key in btree @id at level @level: */
 444 static inline enum btree_node_type __btree_node_type(unsigned level, enum btree_id id)
 445 {
 446         return level ? BKEY_TYPE_BTREE : (enum btree_node_type) id;
 447 }
 448
 449 /* Type of keys @b contains: */
 450 static inline enum btree_node_type btree_node_type(struct btree *b)
 451 {
 452         return __btree_node_type(b->level, b->btree_id);
 453 }
 454
 455 static inline bool btree_node_type_is_extents(enum btree_node_type type)
 456 {
 457         return type == BKEY_TYPE_EXTENTS;
 458 }
 459
 460 static inline bool btree_node_is_extents(struct btree *b)
 461 {
 462         return btree_node_type_is_extents(btree_node_type(b));
 463 }
 464
 465 static inline bool btree_node_type_needs_gc(enum btree_node_type type)
 466 {
 467         switch (type) {
 468         case BKEY_TYPE_ALLOC:
 469         case BKEY_TYPE_BTREE:
 470         case BKEY_TYPE_EXTENTS:
 471         case BKEY_TYPE_INODES:
 472         case BKEY_TYPE_EC:
 473                 return true;
 474         default:
 475                 return false;
 476         }
 477 }
 478
 479 struct btree_root {
 480         struct btree            *b;
 481
 482         struct btree_update     *as;
 483
 484         /* On disk root - see async splits: */
 485         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 486         u8                      level;
 487         u8                      alive;
 488         s8                      error;
 489 };
 490
 491 /*
 492  * Optional hook that will be called just prior to a btree node update, when
 493  * we're holding the write lock and we know what key is about to be overwritten:
 494  */
 495
 496 enum btree_insert_ret {
 497         BTREE_INSERT_OK,
 498         /* leaf node needs to be split */
 499         BTREE_INSERT_BTREE_NODE_FULL,
 500         BTREE_INSERT_ENOSPC,
 501         BTREE_INSERT_NEED_MARK_REPLICAS,
 502         BTREE_INSERT_NEED_JOURNAL_RES,
 503 };
 504
 505 enum btree_gc_coalesce_fail_reason {
 506         BTREE_GC_COALESCE_FAIL_RESERVE_GET,
 507         BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC,
 508         BTREE_GC_COALESCE_FAIL_FORMAT_FITS,
 509 };
 510
 511 enum btree_node_sibling {
 512         btree_prev_sib,
 513         btree_next_sib,
 514 };
 515
 516 typedef struct btree_nr_keys (*sort_fix_overlapping_fn)(struct bset *,
 517                                                         struct btree *,
 518                                                         struct btree_node_iter *);
 519
 520 #endif /* _BCACHEFS_BTREE_TYPES_H */