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 struct btree_trans;
  15
  16 #define MAX_BSETS               3U
  17
  18 struct btree_nr_keys {
  19
  20         /*
  21          * Amount of live metadata (i.e. size of node after a compaction) in
  22          * units of u64s
  23          */
  24         u16                     live_u64s;
  25         u16                     bset_u64s[MAX_BSETS];
  26
  27         /* live keys only: */
  28         u16                     packed_keys;
  29         u16                     unpacked_keys;
  30 };
  31
  32 struct bset_tree {
  33         /*
  34          * We construct a binary tree in an array as if the array
  35          * started at 1, so that things line up on the same cachelines
  36          * better: see comments in bset.c at cacheline_to_bkey() for
  37          * details
  38          */
  39
  40         /* size of the binary tree and prev array */
  41         u16                     size;
  42
  43         /* function of size - precalculated for to_inorder() */
  44         u16                     extra;
  45
  46         u16                     data_offset;
  47         u16                     aux_data_offset;
  48         u16                     end_offset;
  49
  50         struct bpos             max_key;
  51 };
  52
  53 struct btree_write {
  54         struct journal_entry_pin        journal;
  55         struct closure_waitlist         wait;
  56 };
  57
  58 struct btree_alloc {
  59         struct open_buckets     ob;
  60         BKEY_PADDED(k);
  61 };
  62
  63 struct btree_bkey_cached_common {
  64         struct six_lock         lock;
  65         u8                      level;
  66         u8                      btree_id;
  67 };
  68
  69 struct btree {
  70         struct btree_bkey_cached_common c;
  71
  72         struct rhash_head       hash;
  73
  74         /* Key/pointer for this btree node */
  75         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
  76
  77         unsigned long           flags;
  78         u16                     written;
  79         u8                      nsets;
  80         u8                      nr_key_bits;
  81
  82         struct bkey_format      format;
  83
  84         struct btree_node       *data;
  85         void                    *aux_data;
  86
  87         /*
  88          * Sets of sorted keys - the real btree node - plus a binary search tree
  89          *
  90          * set[0] is special; set[0]->tree, set[0]->prev and set[0]->data point
  91          * to the memory we have allocated for this btree node. Additionally,
  92          * set[0]->data points to the entire btree node as it exists on disk.
  93          */
  94         struct bset_tree        set[MAX_BSETS];
  95
  96         struct btree_nr_keys    nr;
  97         u16                     sib_u64s[2];
  98         u16                     whiteout_u64s;
  99         u16                     uncompacted_whiteout_u64s;
 100         u8                      page_order;
 101         u8                      unpack_fn_len;
 102
 103         /*
 104          * XXX: add a delete sequence number, so when bch2_btree_node_relock()
 105          * fails because the lock sequence number has changed - i.e. the
 106          * contents were modified - we can still relock the node if it's still
 107          * the one we want, without redoing the traversal
 108          */
 109
 110         /*
 111          * For asynchronous splits/interior node updates:
 112          * When we do a split, we allocate new child nodes and update the parent
 113          * node to point to them: we update the parent in memory immediately,
 114          * but then we must wait until the children have been written out before
 115          * the update to the parent can be written - this is a list of the
 116          * btree_updates that are blocking this node from being
 117          * written:
 118          */
 119         struct list_head        write_blocked;
 120
 121         /*
 122          * Also for asynchronous splits/interior node updates:
 123          * If a btree node isn't reachable yet, we don't want to kick off
 124          * another write - because that write also won't yet be reachable and
 125          * marking it as completed before it's reachable would be incorrect:
 126          */
 127         unsigned long           will_make_reachable;
 128
 129         struct open_buckets     ob;
 130
 131         /* lru list */
 132         struct list_head        list;
 133
 134         struct btree_write      writes[2];
 135
 136 #ifdef CONFIG_BCACHEFS_DEBUG
 137         bool                    *expensive_debug_checks;
 138 #endif
 139 };
 140
 141 struct btree_cache {
 142         struct rhashtable       table;
 143         bool                    table_init_done;
 144         /*
 145          * We never free a struct btree, except on shutdown - we just put it on
 146          * the btree_cache_freed list and reuse it later. This simplifies the
 147          * code, and it doesn't cost us much memory as the memory usage is
 148          * dominated by buffers that hold the actual btree node data and those
 149          * can be freed - and the number of struct btrees allocated is
 150          * effectively bounded.
 151          *
 152          * btree_cache_freeable effectively is a small cache - we use it because
 153          * high order page allocations can be rather expensive, and it's quite
 154          * common to delete and allocate btree nodes in quick succession. It
 155          * should never grow past ~2-3 nodes in practice.
 156          */
 157         struct mutex            lock;
 158         struct list_head        live;
 159         struct list_head        freeable;
 160         struct list_head        freed;
 161
 162         /* Number of elements in live + freeable lists */
 163         unsigned                used;
 164         unsigned                reserve;
 165         struct shrinker         shrink;
 166
 167         /*
 168          * If we need to allocate memory for a new btree node and that
 169          * allocation fails, we can cannibalize another node in the btree cache
 170          * to satisfy the allocation - lock to guarantee only one thread does
 171          * this at a time:
 172          */
 173         struct task_struct      *alloc_lock;
 174         struct closure_waitlist alloc_wait;
 175 };
 176
 177 struct btree_node_iter {
 178         struct btree_node_iter_set {
 179                 u16     k, end;
 180         } data[MAX_BSETS];
 181 };
 182
 183 enum btree_iter_type {
 184         BTREE_ITER_KEYS,
 185         BTREE_ITER_SLOTS,
 186         BTREE_ITER_NODES,
 187 };
 188
 189 #define BTREE_ITER_TYPE                 ((1 << 2) - 1)
 190
 191 #define BTREE_ITER_INTENT               (1 << 2)
 192 #define BTREE_ITER_PREFETCH             (1 << 3)
 193 /*
 194  * Used in bch2_btree_iter_traverse(), to indicate whether we're searching for
 195  * @pos or the first key strictly greater than @pos
 196  */
 197 #define BTREE_ITER_IS_EXTENTS           (1 << 4)
 198 #define BTREE_ITER_ERROR                (1 << 5)
 199 #define BTREE_ITER_NOUNLOCK             (1 << 6)
 200
 201 enum btree_iter_uptodate {
 202         BTREE_ITER_UPTODATE             = 0,
 203         BTREE_ITER_NEED_PEEK            = 1,
 204         BTREE_ITER_NEED_RELOCK          = 2,
 205         BTREE_ITER_NEED_TRAVERSE        = 3,
 206 };
 207
 208 /*
 209  * @pos                 - iterator's current position
 210  * @level               - current btree depth
 211  * @locks_want          - btree level below which we start taking intent locks
 212  * @nodes_locked        - bitmask indicating which nodes in @nodes are locked
 213  * @nodes_intent_locked - bitmask indicating which locks are intent locks
 214  */
 215 struct btree_iter {
 216         u8                      idx;
 217
 218         struct btree_trans      *trans;
 219         struct bpos             pos;
 220
 221         u8                      flags;
 222         enum btree_iter_uptodate uptodate:4;
 223         enum btree_id           btree_id:4;
 224         unsigned                level:4,
 225                                 locks_want:4,
 226                                 nodes_locked:4,
 227                                 nodes_intent_locked:4;
 228
 229         struct btree_iter_level {
 230                 struct btree    *b;
 231                 struct btree_node_iter iter;
 232                 u32             lock_seq;
 233         }                       l[BTREE_MAX_DEPTH];
 234
 235         /*
 236          * Current unpacked key - so that bch2_btree_iter_next()/
 237          * bch2_btree_iter_next_slot() can correctly advance pos.
 238          */
 239         struct bkey             k;
 240
 241         u64                     id;
 242 };
 243
 244 struct deferred_update {
 245         struct journal_preres   res;
 246         struct journal_entry_pin journal;
 247
 248         spinlock_t              lock;
 249         unsigned                dirty:1;
 250
 251         u8                      allocated_u64s;
 252         enum btree_id           btree_id;
 253
 254         /* must be last: */
 255         struct bkey_i           k;
 256 };
 257
 258 struct btree_insert_entry {
 259         struct bkey_i           *k;
 260
 261         union {
 262         struct btree_iter       *iter;
 263         struct deferred_update  *d;
 264         };
 265
 266         bool                    deferred;
 267 };
 268
 269 #define BTREE_ITER_MAX          64
 270
 271 struct btree_trans {
 272         struct bch_fs           *c;
 273         unsigned long           ip;
 274         size_t                  nr_restarts;
 275         u64                     commit_start;
 276
 277         u64                     iters_linked;
 278         u64                     iters_live;
 279         u64                     iters_touched;
 280         u64                     iters_unlink_on_restart;
 281         u64                     iters_unlink_on_commit;
 282
 283         u8                      nr_iters;
 284         u8                      nr_updates;
 285         u8                      size;
 286         unsigned                used_mempool:1;
 287         unsigned                error:1;
 288
 289         unsigned                mem_top;
 290         unsigned                mem_bytes;
 291         void                    *mem;
 292
 293         struct btree_iter       *iters;
 294         struct btree_insert_entry *updates;
 295
 296         /* update path: */
 297         struct journal_res      journal_res;
 298         struct journal_preres   journal_preres;
 299         u64                     *journal_seq;
 300         struct disk_reservation *disk_res;
 301         unsigned                flags;
 302
 303         struct btree_iter       iters_onstack[2];
 304         struct btree_insert_entry updates_onstack[6];
 305 };
 306
 307 #define BTREE_FLAG(flag)                                                \
 308 static inline bool btree_node_ ## flag(struct btree *b)                 \
 309 {       return test_bit(BTREE_NODE_ ## flag, &b->flags); }              \
 310                                                                         \
 311 static inline void set_btree_node_ ## flag(struct btree *b)             \
 312 {       set_bit(BTREE_NODE_ ## flag, &b->flags); }                      \
 313                                                                         \
 314 static inline void clear_btree_node_ ## flag(struct btree *b)           \
 315 {       clear_bit(BTREE_NODE_ ## flag, &b->flags); }
 316
 317 enum btree_flags {
 318         BTREE_NODE_read_in_flight,
 319         BTREE_NODE_read_error,
 320         BTREE_NODE_dirty,
 321         BTREE_NODE_need_write,
 322         BTREE_NODE_noevict,
 323         BTREE_NODE_write_idx,
 324         BTREE_NODE_accessed,
 325         BTREE_NODE_write_in_flight,
 326         BTREE_NODE_just_written,
 327         BTREE_NODE_dying,
 328         BTREE_NODE_fake,
 329 };
 330
 331 BTREE_FLAG(read_in_flight);
 332 BTREE_FLAG(read_error);
 333 BTREE_FLAG(dirty);
 334 BTREE_FLAG(need_write);
 335 BTREE_FLAG(noevict);
 336 BTREE_FLAG(write_idx);
 337 BTREE_FLAG(accessed);
 338 BTREE_FLAG(write_in_flight);
 339 BTREE_FLAG(just_written);
 340 BTREE_FLAG(dying);
 341 BTREE_FLAG(fake);
 342
 343 static inline struct btree_write *btree_current_write(struct btree *b)
 344 {
 345         return b->writes + btree_node_write_idx(b);
 346 }
 347
 348 static inline struct btree_write *btree_prev_write(struct btree *b)
 349 {
 350         return b->writes + (btree_node_write_idx(b) ^ 1);
 351 }
 352
 353 static inline struct bset_tree *bset_tree_last(struct btree *b)
 354 {
 355         EBUG_ON(!b->nsets);
 356         return b->set + b->nsets - 1;
 357 }
 358
 359 static inline void *
 360 __btree_node_offset_to_ptr(const struct btree *b, u16 offset)
 361 {
 362         return (void *) ((u64 *) b->data + 1 + offset);
 363 }
 364
 365 static inline u16
 366 __btree_node_ptr_to_offset(const struct btree *b, const void *p)
 367 {
 368         u16 ret = (u64 *) p - 1 - (u64 *) b->data;
 369
 370         EBUG_ON(__btree_node_offset_to_ptr(b, ret) != p);
 371         return ret;
 372 }
 373
 374 static inline struct bset *bset(const struct btree *b,
 375                                 const struct bset_tree *t)
 376 {
 377         return __btree_node_offset_to_ptr(b, t->data_offset);
 378 }
 379
 380 static inline void set_btree_bset_end(struct btree *b, struct bset_tree *t)
 381 {
 382         t->end_offset =
 383                 __btree_node_ptr_to_offset(b, vstruct_last(bset(b, t)));
 384 }
 385
 386 static inline void set_btree_bset(struct btree *b, struct bset_tree *t,
 387                                   const struct bset *i)
 388 {
 389         t->data_offset = __btree_node_ptr_to_offset(b, i);
 390         set_btree_bset_end(b, t);
 391 }
 392
 393 static inline struct bset *btree_bset_first(struct btree *b)
 394 {
 395         return bset(b, b->set);
 396 }
 397
 398 static inline struct bset *btree_bset_last(struct btree *b)
 399 {
 400         return bset(b, bset_tree_last(b));
 401 }
 402
 403 static inline u16
 404 __btree_node_key_to_offset(const struct btree *b, const struct bkey_packed *k)
 405 {
 406         return __btree_node_ptr_to_offset(b, k);
 407 }
 408
 409 static inline struct bkey_packed *
 410 __btree_node_offset_to_key(const struct btree *b, u16 k)
 411 {
 412         return __btree_node_offset_to_ptr(b, k);
 413 }
 414
 415 static inline unsigned btree_bkey_first_offset(const struct bset_tree *t)
 416 {
 417         return t->data_offset + offsetof(struct bset, _data) / sizeof(u64);
 418 }
 419
 420 #define btree_bkey_first(_b, _t)                                        \
 421 ({                                                                      \
 422         EBUG_ON(bset(_b, _t)->start !=                                  \
 423                 __btree_node_offset_to_key(_b, btree_bkey_first_offset(_t)));\
 424                                                                         \
 425         bset(_b, _t)->start;                                            \
 426 })
 427
 428 #define btree_bkey_last(_b, _t)                                         \
 429 ({                                                                      \
 430         EBUG_ON(__btree_node_offset_to_key(_b, (_t)->end_offset) !=     \
 431                 vstruct_last(bset(_b, _t)));                            \
 432                                                                         \
 433         __btree_node_offset_to_key(_b, (_t)->end_offset);               \
 434 })
 435
 436 static inline unsigned bset_byte_offset(struct btree *b, void *i)
 437 {
 438         return i - (void *) b->data;
 439 }
 440
 441 enum btree_node_type {
 442 #define x(kwd, val, name) BKEY_TYPE_##kwd = val,
 443         BCH_BTREE_IDS()
 444 #undef x
 445         BKEY_TYPE_BTREE,
 446 };
 447
 448 /* Type of a key in btree @id at level @level: */
 449 static inline enum btree_node_type __btree_node_type(unsigned level, enum btree_id id)
 450 {
 451         return level ? BKEY_TYPE_BTREE : (enum btree_node_type) id;
 452 }
 453
 454 /* Type of keys @b contains: */
 455 static inline enum btree_node_type btree_node_type(struct btree *b)
 456 {
 457         return __btree_node_type(b->c.level, b->c.btree_id);
 458 }
 459
 460 static inline bool btree_node_type_is_extents(enum btree_node_type type)
 461 {
 462         return type == BKEY_TYPE_EXTENTS;
 463 }
 464
 465 static inline bool btree_node_is_extents(struct btree *b)
 466 {
 467         return btree_node_type_is_extents(btree_node_type(b));
 468 }
 469
 470 static inline bool btree_node_type_needs_gc(enum btree_node_type type)
 471 {
 472         switch (type) {
 473         case BKEY_TYPE_ALLOC:
 474         case BKEY_TYPE_BTREE:
 475         case BKEY_TYPE_EXTENTS:
 476         case BKEY_TYPE_INODES:
 477         case BKEY_TYPE_EC:
 478                 return true;
 479         default:
 480                 return false;
 481         }
 482 }
 483
 484 struct btree_root {
 485         struct btree            *b;
 486
 487         struct btree_update     *as;
 488
 489         /* On disk root - see async splits: */
 490         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 491         u8                      level;
 492         u8                      alive;
 493         s8                      error;
 494 };
 495
 496 /*
 497  * Optional hook that will be called just prior to a btree node update, when
 498  * we're holding the write lock and we know what key is about to be overwritten:
 499  */
 500
 501 enum btree_insert_ret {
 502         BTREE_INSERT_OK,
 503         /* leaf node needs to be split */
 504         BTREE_INSERT_BTREE_NODE_FULL,
 505         BTREE_INSERT_ENOSPC,
 506         BTREE_INSERT_NEED_MARK_REPLICAS,
 507         BTREE_INSERT_NEED_JOURNAL_RES,
 508 };
 509
 510 enum btree_gc_coalesce_fail_reason {
 511         BTREE_GC_COALESCE_FAIL_RESERVE_GET,
 512         BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC,
 513         BTREE_GC_COALESCE_FAIL_FORMAT_FITS,
 514 };
 515
 516 enum btree_node_sibling {
 517         btree_prev_sib,
 518         btree_next_sib,
 519 };
 520
 521 typedef struct btree_nr_keys (*sort_fix_overlapping_fn)(struct bset *,
 522                                                         struct btree *,
 523                                                         struct btree_node_iter *);
 524
 525 #endif /* _BCACHEFS_BTREE_TYPES_H */