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