fs/bcachefs/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
   8 #include "bkey_methods.h"
   9 #include "journal_types.h"
  10 #include "six.h"
  11
  12 struct open_bucket;
  13 struct btree_update;
  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
 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 btree_insert_entry {
 248         struct btree_iter *iter;
 249         struct bkey_i   *k;
 250 };
 251
 252 struct btree_trans {
 253         struct bch_fs           *c;
 254         size_t                  nr_restarts;
 255
 256         u8                      nr_iters;
 257         u8                      iters_live;
 258         u8                      iters_linked;
 259         u8                      nr_updates;
 260
 261         unsigned                mem_top;
 262         unsigned                mem_bytes;
 263         void                    *mem;
 264
 265         struct btree_iter       *iters;
 266         u64                     iter_ids[BTREE_ITER_MAX];
 267
 268         struct btree_insert_entry updates[BTREE_ITER_MAX];
 269
 270         struct btree_iter       iters_onstack[2];
 271 };
 272
 273 #define BTREE_FLAG(flag)                                                \
 274 static inline bool btree_node_ ## flag(struct btree *b)                 \
 275 {       return test_bit(BTREE_NODE_ ## flag, &b->flags); }              \
 276                                                                         \
 277 static inline void set_btree_node_ ## flag(struct btree *b)             \
 278 {       set_bit(BTREE_NODE_ ## flag, &b->flags); }                      \
 279                                                                         \
 280 static inline void clear_btree_node_ ## flag(struct btree *b)           \
 281 {       clear_bit(BTREE_NODE_ ## flag, &b->flags); }
 282
 283 enum btree_flags {
 284         BTREE_NODE_read_in_flight,
 285         BTREE_NODE_read_error,
 286         BTREE_NODE_dirty,
 287         BTREE_NODE_need_write,
 288         BTREE_NODE_noevict,
 289         BTREE_NODE_write_idx,
 290         BTREE_NODE_accessed,
 291         BTREE_NODE_write_in_flight,
 292         BTREE_NODE_just_written,
 293         BTREE_NODE_dying,
 294         BTREE_NODE_fake,
 295 };
 296
 297 BTREE_FLAG(read_in_flight);
 298 BTREE_FLAG(read_error);
 299 BTREE_FLAG(dirty);
 300 BTREE_FLAG(need_write);
 301 BTREE_FLAG(noevict);
 302 BTREE_FLAG(write_idx);
 303 BTREE_FLAG(accessed);
 304 BTREE_FLAG(write_in_flight);
 305 BTREE_FLAG(just_written);
 306 BTREE_FLAG(dying);
 307 BTREE_FLAG(fake);
 308
 309 static inline struct btree_write *btree_current_write(struct btree *b)
 310 {
 311         return b->writes + btree_node_write_idx(b);
 312 }
 313
 314 static inline struct btree_write *btree_prev_write(struct btree *b)
 315 {
 316         return b->writes + (btree_node_write_idx(b) ^ 1);
 317 }
 318
 319 static inline struct bset_tree *bset_tree_last(struct btree *b)
 320 {
 321         EBUG_ON(!b->nsets);
 322         return b->set + b->nsets - 1;
 323 }
 324
 325 static inline void *
 326 __btree_node_offset_to_ptr(const struct btree *b, u16 offset)
 327 {
 328         return (void *) ((u64 *) b->data + 1 + offset);
 329 }
 330
 331 static inline u16
 332 __btree_node_ptr_to_offset(const struct btree *b, const void *p)
 333 {
 334         u16 ret = (u64 *) p - 1 - (u64 *) b->data;
 335
 336         EBUG_ON(__btree_node_offset_to_ptr(b, ret) != p);
 337         return ret;
 338 }
 339
 340 static inline struct bset *bset(const struct btree *b,
 341                                 const struct bset_tree *t)
 342 {
 343         return __btree_node_offset_to_ptr(b, t->data_offset);
 344 }
 345
 346 static inline void set_btree_bset_end(struct btree *b, struct bset_tree *t)
 347 {
 348         t->end_offset =
 349                 __btree_node_ptr_to_offset(b, vstruct_last(bset(b, t)));
 350 }
 351
 352 static inline void set_btree_bset(struct btree *b, struct bset_tree *t,
 353                                   const struct bset *i)
 354 {
 355         t->data_offset = __btree_node_ptr_to_offset(b, i);
 356         set_btree_bset_end(b, t);
 357 }
 358
 359 static inline struct bset *btree_bset_first(struct btree *b)
 360 {
 361         return bset(b, b->set);
 362 }
 363
 364 static inline struct bset *btree_bset_last(struct btree *b)
 365 {
 366         return bset(b, bset_tree_last(b));
 367 }
 368
 369 static inline u16
 370 __btree_node_key_to_offset(const struct btree *b, const struct bkey_packed *k)
 371 {
 372         return __btree_node_ptr_to_offset(b, k);
 373 }
 374
 375 static inline struct bkey_packed *
 376 __btree_node_offset_to_key(const struct btree *b, u16 k)
 377 {
 378         return __btree_node_offset_to_ptr(b, k);
 379 }
 380
 381 static inline unsigned btree_bkey_first_offset(const struct bset_tree *t)
 382 {
 383         return t->data_offset + offsetof(struct bset, _data) / sizeof(u64);
 384 }
 385
 386 #define btree_bkey_first(_b, _t)                                        \
 387 ({                                                                      \
 388         EBUG_ON(bset(_b, _t)->start !=                                  \
 389                 __btree_node_offset_to_key(_b, btree_bkey_first_offset(_t)));\
 390                                                                         \
 391         bset(_b, _t)->start;                                            \
 392 })
 393
 394 #define btree_bkey_last(_b, _t)                                         \
 395 ({                                                                      \
 396         EBUG_ON(__btree_node_offset_to_key(_b, (_t)->end_offset) !=     \
 397                 vstruct_last(bset(_b, _t)));                            \
 398                                                                         \
 399         __btree_node_offset_to_key(_b, (_t)->end_offset);               \
 400 })
 401
 402 static inline unsigned bset_byte_offset(struct btree *b, void *i)
 403 {
 404         return i - (void *) b->data;
 405 }
 406
 407 /* Type of keys @b contains: */
 408 static inline enum bkey_type btree_node_type(struct btree *b)
 409 {
 410         return b->level ? BKEY_TYPE_BTREE : b->btree_id;
 411 }
 412
 413 static inline const struct bkey_ops *btree_node_ops(struct btree *b)
 414 {
 415         return &bch2_bkey_ops[btree_node_type(b)];
 416 }
 417
 418 static inline bool btree_node_has_ptrs(struct btree *b)
 419 {
 420         return btree_type_has_ptrs(btree_node_type(b));
 421 }
 422
 423 static inline bool btree_node_is_extents(struct btree *b)
 424 {
 425         return btree_node_type(b) == BKEY_TYPE_EXTENTS;
 426 }
 427
 428 struct btree_root {
 429         struct btree            *b;
 430
 431         struct btree_update     *as;
 432
 433         /* On disk root - see async splits: */
 434         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 435         u8                      level;
 436         u8                      alive;
 437 };
 438
 439 /*
 440  * Optional hook that will be called just prior to a btree node update, when
 441  * we're holding the write lock and we know what key is about to be overwritten:
 442  */
 443
 444 enum btree_insert_ret {
 445         BTREE_INSERT_OK,
 446         /* extent spanned multiple leaf nodes: have to traverse to next node: */
 447         BTREE_INSERT_NEED_TRAVERSE,
 448         /* write lock held for too long */
 449         /* leaf node needs to be split */
 450         BTREE_INSERT_BTREE_NODE_FULL,
 451         BTREE_INSERT_ENOSPC,
 452         BTREE_INSERT_NEED_GC_LOCK,
 453 };
 454
 455 enum btree_gc_coalesce_fail_reason {
 456         BTREE_GC_COALESCE_FAIL_RESERVE_GET,
 457         BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC,
 458         BTREE_GC_COALESCE_FAIL_FORMAT_FITS,
 459 };
 460
 461 enum btree_node_sibling {
 462         btree_prev_sib,
 463         btree_next_sib,
 464 };
 465
 466 typedef struct btree_nr_keys (*sort_fix_overlapping_fn)(struct bset *,
 467                                                         struct btree *,
 468                                                         struct btree_node_iter *);
 469
 470 #endif /* _BCACHEFS_BTREE_TYPES_H */