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
 195 enum btree_iter_uptodate {
 196         BTREE_ITER_UPTODATE             = 0,
 197         BTREE_ITER_NEED_PEEK            = 1,
 198         BTREE_ITER_NEED_RELOCK          = 2,
 199         BTREE_ITER_NEED_TRAVERSE        = 3,
 200 };
 201
 202 /*
 203  * @pos                 - iterator's current position
 204  * @level               - current btree depth
 205  * @locks_want          - btree level below which we start taking intent locks
 206  * @nodes_locked        - bitmask indicating which nodes in @nodes are locked
 207  * @nodes_intent_locked - bitmask indicating which locks are intent locks
 208  */
 209 struct btree_iter {
 210         struct bch_fs           *c;
 211         struct bpos             pos;
 212
 213         u8                      flags;
 214         enum btree_iter_uptodate uptodate:4;
 215         enum btree_id           btree_id:4;
 216         unsigned                level:4,
 217                                 locks_want:4,
 218                                 nodes_locked:4,
 219                                 nodes_intent_locked:4;
 220
 221         struct btree_iter_level {
 222                 struct btree    *b;
 223                 struct btree_node_iter iter;
 224                 u32             lock_seq;
 225         }                       l[BTREE_MAX_DEPTH];
 226
 227         /*
 228          * Current unpacked key - so that bch2_btree_iter_next()/
 229          * bch2_btree_iter_next_slot() can correctly advance pos.
 230          */
 231         struct bkey             k;
 232
 233         /*
 234          * Circular linked list of linked iterators: linked iterators share
 235          * locks (e.g. two linked iterators may have the same node intent
 236          * locked, or read and write locked, at the same time), and insertions
 237          * through one iterator won't invalidate the other linked iterators.
 238          */
 239
 240         /* Must come last: */
 241         struct btree_iter       *next;
 242 };
 243
 244 #define BTREE_ITER_MAX          8
 245
 246 struct btree_insert_entry {
 247         struct btree_iter *iter;
 248         struct bkey_i   *k;
 249 };
 250
 251 struct btree_trans {
 252         struct bch_fs           *c;
 253         size_t                  nr_restarts;
 254
 255         u8                      nr_iters;
 256         u8                      iters_live;
 257         u8                      iters_linked;
 258         u8                      nr_updates;
 259
 260         unsigned                mem_top;
 261         unsigned                mem_bytes;
 262         void                    *mem;
 263
 264         struct btree_iter       *iters;
 265         u64                     iter_ids[BTREE_ITER_MAX];
 266
 267         struct btree_insert_entry updates[BTREE_ITER_MAX];
 268
 269         struct btree_iter       iters_onstack[2];
 270 };
 271
 272 #define BTREE_FLAG(flag)                                                \
 273 static inline bool btree_node_ ## flag(struct btree *b)                 \
 274 {       return test_bit(BTREE_NODE_ ## flag, &b->flags); }              \
 275                                                                         \
 276 static inline void set_btree_node_ ## flag(struct btree *b)             \
 277 {       set_bit(BTREE_NODE_ ## flag, &b->flags); }                      \
 278                                                                         \
 279 static inline void clear_btree_node_ ## flag(struct btree *b)           \
 280 {       clear_bit(BTREE_NODE_ ## flag, &b->flags); }
 281
 282 enum btree_flags {
 283         BTREE_NODE_read_in_flight,
 284         BTREE_NODE_read_error,
 285         BTREE_NODE_dirty,
 286         BTREE_NODE_need_write,
 287         BTREE_NODE_noevict,
 288         BTREE_NODE_write_idx,
 289         BTREE_NODE_accessed,
 290         BTREE_NODE_write_in_flight,
 291         BTREE_NODE_just_written,
 292         BTREE_NODE_dying,
 293         BTREE_NODE_fake,
 294 };
 295
 296 BTREE_FLAG(read_in_flight);
 297 BTREE_FLAG(read_error);
 298 BTREE_FLAG(dirty);
 299 BTREE_FLAG(need_write);
 300 BTREE_FLAG(noevict);
 301 BTREE_FLAG(write_idx);
 302 BTREE_FLAG(accessed);
 303 BTREE_FLAG(write_in_flight);
 304 BTREE_FLAG(just_written);
 305 BTREE_FLAG(dying);
 306 BTREE_FLAG(fake);
 307
 308 static inline struct btree_write *btree_current_write(struct btree *b)
 309 {
 310         return b->writes + btree_node_write_idx(b);
 311 }
 312
 313 static inline struct btree_write *btree_prev_write(struct btree *b)
 314 {
 315         return b->writes + (btree_node_write_idx(b) ^ 1);
 316 }
 317
 318 static inline struct bset_tree *bset_tree_last(struct btree *b)
 319 {
 320         EBUG_ON(!b->nsets);
 321         return b->set + b->nsets - 1;
 322 }
 323
 324 static inline void *
 325 __btree_node_offset_to_ptr(const struct btree *b, u16 offset)
 326 {
 327         return (void *) ((u64 *) b->data + 1 + offset);
 328 }
 329
 330 static inline u16
 331 __btree_node_ptr_to_offset(const struct btree *b, const void *p)
 332 {
 333         u16 ret = (u64 *) p - 1 - (u64 *) b->data;
 334
 335         EBUG_ON(__btree_node_offset_to_ptr(b, ret) != p);
 336         return ret;
 337 }
 338
 339 static inline struct bset *bset(const struct btree *b,
 340                                 const struct bset_tree *t)
 341 {
 342         return __btree_node_offset_to_ptr(b, t->data_offset);
 343 }
 344
 345 static inline void set_btree_bset_end(struct btree *b, struct bset_tree *t)
 346 {
 347         t->end_offset =
 348                 __btree_node_ptr_to_offset(b, vstruct_last(bset(b, t)));
 349 }
 350
 351 static inline void set_btree_bset(struct btree *b, struct bset_tree *t,
 352                                   const struct bset *i)
 353 {
 354         t->data_offset = __btree_node_ptr_to_offset(b, i);
 355         set_btree_bset_end(b, t);
 356 }
 357
 358 static inline struct bset *btree_bset_first(struct btree *b)
 359 {
 360         return bset(b, b->set);
 361 }
 362
 363 static inline struct bset *btree_bset_last(struct btree *b)
 364 {
 365         return bset(b, bset_tree_last(b));
 366 }
 367
 368 static inline u16
 369 __btree_node_key_to_offset(const struct btree *b, const struct bkey_packed *k)
 370 {
 371         return __btree_node_ptr_to_offset(b, k);
 372 }
 373
 374 static inline struct bkey_packed *
 375 __btree_node_offset_to_key(const struct btree *b, u16 k)
 376 {
 377         return __btree_node_offset_to_ptr(b, k);
 378 }
 379
 380 static inline unsigned btree_bkey_first_offset(const struct bset_tree *t)
 381 {
 382         return t->data_offset + offsetof(struct bset, _data) / sizeof(u64);
 383 }
 384
 385 #define btree_bkey_first(_b, _t)                                        \
 386 ({                                                                      \
 387         EBUG_ON(bset(_b, _t)->start !=                                  \
 388                 __btree_node_offset_to_key(_b, btree_bkey_first_offset(_t)));\
 389                                                                         \
 390         bset(_b, _t)->start;                                            \
 391 })
 392
 393 #define btree_bkey_last(_b, _t)                                         \
 394 ({                                                                      \
 395         EBUG_ON(__btree_node_offset_to_key(_b, (_t)->end_offset) !=     \
 396                 vstruct_last(bset(_b, _t)));                            \
 397                                                                         \
 398         __btree_node_offset_to_key(_b, (_t)->end_offset);               \
 399 })
 400
 401 static inline unsigned bset_byte_offset(struct btree *b, void *i)
 402 {
 403         return i - (void *) b->data;
 404 }
 405
 406 /* Type of keys @b contains: */
 407 static inline enum bkey_type btree_node_type(struct btree *b)
 408 {
 409         return b->level ? BKEY_TYPE_BTREE : b->btree_id;
 410 }
 411
 412 static inline const struct bkey_ops *btree_node_ops(struct btree *b)
 413 {
 414         return &bch2_bkey_ops[btree_node_type(b)];
 415 }
 416
 417 static inline bool btree_node_is_extents(struct btree *b)
 418 {
 419         return btree_node_type(b) == BKEY_TYPE_EXTENTS;
 420 }
 421
 422 struct btree_root {
 423         struct btree            *b;
 424
 425         struct btree_update     *as;
 426
 427         /* On disk root - see async splits: */
 428         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 429         u8                      level;
 430         u8                      alive;
 431 };
 432
 433 /*
 434  * Optional hook that will be called just prior to a btree node update, when
 435  * we're holding the write lock and we know what key is about to be overwritten:
 436  */
 437
 438 enum btree_insert_ret {
 439         BTREE_INSERT_OK,
 440         /* extent spanned multiple leaf nodes: have to traverse to next node: */
 441         BTREE_INSERT_NEED_TRAVERSE,
 442         /* write lock held for too long */
 443         /* leaf node needs to be split */
 444         BTREE_INSERT_BTREE_NODE_FULL,
 445         BTREE_INSERT_ENOSPC,
 446         BTREE_INSERT_NEED_GC_LOCK,
 447 };
 448
 449 enum btree_gc_coalesce_fail_reason {
 450         BTREE_GC_COALESCE_FAIL_RESERVE_GET,
 451         BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC,
 452         BTREE_GC_COALESCE_FAIL_FORMAT_FITS,
 453 };
 454
 455 enum btree_node_sibling {
 456         btree_prev_sib,
 457         btree_next_sib,
 458 };
 459
 460 typedef struct btree_nr_keys (*sort_fix_overlapping_fn)(struct bset *,
 461                                                         struct btree *,
 462                                                         struct btree_node_iter *);
 463
 464 #endif /* _BCACHEFS_BTREE_TYPES_H */