git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/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 #include <linux/six.h>
   8
   9 #include "bkey_methods.h"
  10 #include "buckets_types.h"
  11 #include "journal_types.h"
  12
  13 struct open_bucket;
  14 struct btree_update;
  15 struct btree_trans;
  16
  17 #define MAX_BSETS               3U
  18
  19 struct btree_nr_keys {
  20
  21         /*
  22          * Amount of live metadata (i.e. size of node after a compaction) in
  23          * units of u64s
  24          */
  25         u16                     live_u64s;
  26         u16                     bset_u64s[MAX_BSETS];
  27
  28         /* live keys only: */
  29         u16                     packed_keys;
  30         u16                     unpacked_keys;
  31 };
  32
  33 struct bset_tree {
  34         /*
  35          * We construct a binary tree in an array as if the array
  36          * started at 1, so that things line up on the same cachelines
  37          * better: see comments in bset.c at cacheline_to_bkey() for
  38          * details
  39          */
  40
  41         /* size of the binary tree and prev array */
  42         u16                     size;
  43
  44         /* function of size - precalculated for to_inorder() */
  45         u16                     extra;
  46
  47         u16                     data_offset;
  48         u16                     aux_data_offset;
  49         u16                     end_offset;
  50
  51         struct bpos             max_key;
  52 };
  53
  54 struct btree_write {
  55         struct journal_entry_pin        journal;
  56         struct closure_waitlist         wait;
  57 };
  58
  59 struct btree_alloc {
  60         struct open_buckets     ob;
  61         BKEY_PADDED(k);
  62 };
  63
  64 struct btree {
  65         /* Hottest entries first */
  66         struct rhash_head       hash;
  67
  68         /* Key/pointer for this btree node */
  69         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
  70
  71         struct six_lock         lock;
  72
  73         unsigned long           flags;
  74         u16                     written;
  75         u8                      level;
  76         u8                      btree_id;
  77         u8                      nsets;
  78         u8                      nr_key_bits;
  79
  80         struct bkey_format      format;
  81
  82         struct btree_node       *data;
  83         void                    *aux_data;
  84
  85         /*
  86          * Sets of sorted keys - the real btree node - plus a binary search tree
  87          *
  88          * set[0] is special; set[0]->tree, set[0]->prev and set[0]->data point
  89          * to the memory we have allocated for this btree node. Additionally,
  90          * set[0]->data points to the entire btree node as it exists on disk.
  91          */
  92         struct bset_tree        set[MAX_BSETS];
  93
  94         struct btree_nr_keys    nr;
  95         u16                     sib_u64s[2];
  96         u16                     whiteout_u64s;
  97         u8                      page_order;
  98         u8                      unpack_fn_len;
  99
 100         /*
 101          * XXX: add a delete sequence number, so when bch2_btree_node_relock()
 102          * fails because the lock sequence number has changed - i.e. the
 103          * contents were modified - we can still relock the node if it's still
 104          * the one we want, without redoing the traversal
 105          */
 106
 107         /*
 108          * For asynchronous splits/interior node updates:
 109          * When we do a split, we allocate new child nodes and update the parent
 110          * node to point to them: we update the parent in memory immediately,
 111          * but then we must wait until the children have been written out before
 112          * the update to the parent can be written - this is a list of the
 113          * btree_updates that are blocking this node from being
 114          * written:
 115          */
 116         struct list_head        write_blocked;
 117
 118         /*
 119          * Also for asynchronous splits/interior node updates:
 120          * If a btree node isn't reachable yet, we don't want to kick off
 121          * another write - because that write also won't yet be reachable and
 122          * marking it as completed before it's reachable would be incorrect:
 123          */
 124         unsigned long           will_make_reachable;
 125
 126         struct open_buckets     ob;
 127
 128         /* lru list */
 129         struct list_head        list;
 130
 131         struct btree_write      writes[2];
 132
 133 #ifdef CONFIG_BCACHEFS_DEBUG
 134         bool                    *expensive_debug_checks;
 135 #endif
 136 };
 137
 138 struct btree_cache {
 139         struct rhashtable       table;
 140         bool                    table_init_done;
 141         /*
 142          * We never free a struct btree, except on shutdown - we just put it on
 143          * the btree_cache_freed list and reuse it later. This simplifies the
 144          * code, and it doesn't cost us much memory as the memory usage is
 145          * dominated by buffers that hold the actual btree node data and those
 146          * can be freed - and the number of struct btrees allocated is
 147          * effectively bounded.
 148          *
 149          * btree_cache_freeable effectively is a small cache - we use it because
 150          * high order page allocations can be rather expensive, and it's quite
 151          * common to delete and allocate btree nodes in quick succession. It
 152          * should never grow past ~2-3 nodes in practice.
 153          */
 154         struct mutex            lock;
 155         struct list_head        live;
 156         struct list_head        freeable;
 157         struct list_head        freed;
 158
 159         /* Number of elements in live + freeable lists */
 160         unsigned                used;
 161         unsigned                reserve;
 162         struct shrinker         shrink;
 163
 164         /*
 165          * If we need to allocate memory for a new btree node and that
 166          * allocation fails, we can cannibalize another node in the btree cache
 167          * to satisfy the allocation - lock to guarantee only one thread does
 168          * this at a time:
 169          */
 170         struct task_struct      *alloc_lock;
 171         struct closure_waitlist alloc_wait;
 172 };
 173
 174 struct btree_node_iter {
 175         struct btree_node_iter_set {
 176                 u16     k, end;
 177         } data[MAX_BSETS];
 178 };
 179
 180 enum btree_iter_type {
 181         BTREE_ITER_KEYS,
 182         BTREE_ITER_NODES,
 183 };
 184
 185 #define BTREE_ITER_TYPE                 ((1 << 2) - 1)
 186
 187 #define BTREE_ITER_SLOTS                (1 << 2)
 188 #define BTREE_ITER_INTENT               (1 << 3)
 189 #define BTREE_ITER_PREFETCH             (1 << 4)
 190 #define BTREE_ITER_KEEP_UNTIL_COMMIT    (1 << 5)
 191 /*
 192  * Used in bch2_btree_iter_traverse(), to indicate whether we're searching for
 193  * @pos or the first key strictly greater than @pos
 194  */
 195 #define BTREE_ITER_IS_EXTENTS           (1 << 6)
 196 #define BTREE_ITER_ERROR                (1 << 7)
 197
 198 enum btree_iter_uptodate {
 199         BTREE_ITER_UPTODATE             = 0,
 200         BTREE_ITER_NEED_PEEK            = 1,
 201         BTREE_ITER_NEED_RELOCK          = 2,
 202         BTREE_ITER_NEED_TRAVERSE        = 3,
 203 };
 204
 205 /*
 206  * @pos                 - iterator's current position
 207  * @level               - current btree depth
 208  * @locks_want          - btree level below which we start taking intent locks
 209  * @nodes_locked        - bitmask indicating which nodes in @nodes are locked
 210  * @nodes_intent_locked - bitmask indicating which locks are intent locks
 211  */
 212 struct btree_iter {
 213         u8                      idx;
 214
 215         struct btree_trans      *trans;
 216         struct bpos             pos;
 217
 218         u8                      flags;
 219         enum btree_iter_uptodate uptodate:4;
 220         enum btree_id           btree_id:4;
 221         unsigned                level:4,
 222                                 locks_want:4,
 223                                 nodes_locked:4,
 224                                 nodes_intent_locked:4;
 225
 226         struct btree_iter_level {
 227                 struct btree    *b;
 228                 struct btree_node_iter iter;
 229                 u32             lock_seq;
 230         }                       l[BTREE_MAX_DEPTH];
 231
 232         /*
 233          * Current unpacked key - so that bch2_btree_iter_next()/
 234          * bch2_btree_iter_next_slot() can correctly advance pos.
 235          */
 236         struct bkey             k;
 237 };
 238
 239 static inline enum btree_iter_type btree_iter_type(struct btree_iter *iter)
 240 {
 241         return iter->flags & BTREE_ITER_TYPE;
 242 }
 243
 244 struct btree_insert_entry {
 245         struct bkey_i           *k;
 246         struct btree_iter       *iter;
 247 };
 248
 249 #define BTREE_ITER_MAX          64
 250
 251 struct btree_trans {
 252         struct bch_fs           *c;
 253         unsigned long           ip;
 254
 255         u64                     iters_linked;
 256         u64                     iters_live;
 257         u64                     iters_touched;
 258
 259         u8                      nr_iters;
 260         u8                      nr_updates;
 261         u8                      size;
 262         unsigned                used_mempool:1;
 263         unsigned                error:1;
 264         unsigned                nounlock:1;
 265
 266         unsigned                mem_top;
 267         unsigned                mem_bytes;
 268         void                    *mem;
 269
 270         struct btree_iter       *iters;
 271         struct btree_insert_entry *updates;
 272         u8                      *updates_sorted;
 273
 274         /* update path: */
 275         struct journal_res      journal_res;
 276         struct journal_preres   journal_preres;
 277         u64                     *journal_seq;
 278         struct disk_reservation *disk_res;
 279         unsigned                flags;
 280         unsigned                journal_u64s;
 281         unsigned                journal_preres_u64s;
 282         struct replicas_delta_list *fs_usage_deltas;
 283
 284         struct btree_iter       iters_onstack[2];
 285         struct btree_insert_entry updates_onstack[6];
 286         u8                      updates_sorted_onstack[6];
 287 };
 288
 289 #define BTREE_FLAG(flag)                                                \
 290 static inline bool btree_node_ ## flag(struct btree *b)                 \
 291 {       return test_bit(BTREE_NODE_ ## flag, &b->flags); }              \
 292                                                                         \
 293 static inline void set_btree_node_ ## flag(struct btree *b)             \
 294 {       set_bit(BTREE_NODE_ ## flag, &b->flags); }                      \
 295                                                                         \
 296 static inline void clear_btree_node_ ## flag(struct btree *b)           \
 297 {       clear_bit(BTREE_NODE_ ## flag, &b->flags); }
 298
 299 enum btree_flags {
 300         BTREE_NODE_read_in_flight,
 301         BTREE_NODE_read_error,
 302         BTREE_NODE_dirty,
 303         BTREE_NODE_need_write,
 304         BTREE_NODE_noevict,
 305         BTREE_NODE_write_idx,
 306         BTREE_NODE_accessed,
 307         BTREE_NODE_write_in_flight,
 308         BTREE_NODE_just_written,
 309         BTREE_NODE_dying,
 310         BTREE_NODE_fake,
 311 };
 312
 313 BTREE_FLAG(read_in_flight);
 314 BTREE_FLAG(read_error);
 315 BTREE_FLAG(dirty);
 316 BTREE_FLAG(need_write);
 317 BTREE_FLAG(noevict);
 318 BTREE_FLAG(write_idx);
 319 BTREE_FLAG(accessed);
 320 BTREE_FLAG(write_in_flight);
 321 BTREE_FLAG(just_written);
 322 BTREE_FLAG(dying);
 323 BTREE_FLAG(fake);
 324
 325 static inline struct btree_write *btree_current_write(struct btree *b)
 326 {
 327         return b->writes + btree_node_write_idx(b);
 328 }
 329
 330 static inline struct btree_write *btree_prev_write(struct btree *b)
 331 {
 332         return b->writes + (btree_node_write_idx(b) ^ 1);
 333 }
 334
 335 static inline struct bset_tree *bset_tree_last(struct btree *b)
 336 {
 337         EBUG_ON(!b->nsets);
 338         return b->set + b->nsets - 1;
 339 }
 340
 341 static inline void *
 342 __btree_node_offset_to_ptr(const struct btree *b, u16 offset)
 343 {
 344         return (void *) ((u64 *) b->data + 1 + offset);
 345 }
 346
 347 static inline u16
 348 __btree_node_ptr_to_offset(const struct btree *b, const void *p)
 349 {
 350         u16 ret = (u64 *) p - 1 - (u64 *) b->data;
 351
 352         EBUG_ON(__btree_node_offset_to_ptr(b, ret) != p);
 353         return ret;
 354 }
 355
 356 static inline struct bset *bset(const struct btree *b,
 357                                 const struct bset_tree *t)
 358 {
 359         return __btree_node_offset_to_ptr(b, t->data_offset);
 360 }
 361
 362 static inline void set_btree_bset_end(struct btree *b, struct bset_tree *t)
 363 {
 364         t->end_offset =
 365                 __btree_node_ptr_to_offset(b, vstruct_last(bset(b, t)));
 366 }
 367
 368 static inline void set_btree_bset(struct btree *b, struct bset_tree *t,
 369                                   const struct bset *i)
 370 {
 371         t->data_offset = __btree_node_ptr_to_offset(b, i);
 372         set_btree_bset_end(b, t);
 373 }
 374
 375 static inline struct bset *btree_bset_first(struct btree *b)
 376 {
 377         return bset(b, b->set);
 378 }
 379
 380 static inline struct bset *btree_bset_last(struct btree *b)
 381 {
 382         return bset(b, bset_tree_last(b));
 383 }
 384
 385 static inline u16
 386 __btree_node_key_to_offset(const struct btree *b, const struct bkey_packed *k)
 387 {
 388         return __btree_node_ptr_to_offset(b, k);
 389 }
 390
 391 static inline struct bkey_packed *
 392 __btree_node_offset_to_key(const struct btree *b, u16 k)
 393 {
 394         return __btree_node_offset_to_ptr(b, k);
 395 }
 396
 397 static inline unsigned btree_bkey_first_offset(const struct bset_tree *t)
 398 {
 399         return t->data_offset + offsetof(struct bset, _data) / sizeof(u64);
 400 }
 401
 402 #define btree_bkey_first(_b, _t)                                        \
 403 ({                                                                      \
 404         EBUG_ON(bset(_b, _t)->start !=                                  \
 405                 __btree_node_offset_to_key(_b, btree_bkey_first_offset(_t)));\
 406                                                                         \
 407         bset(_b, _t)->start;                                            \
 408 })
 409
 410 #define btree_bkey_last(_b, _t)                                         \
 411 ({                                                                      \
 412         EBUG_ON(__btree_node_offset_to_key(_b, (_t)->end_offset) !=     \
 413                 vstruct_last(bset(_b, _t)));                            \
 414                                                                         \
 415         __btree_node_offset_to_key(_b, (_t)->end_offset);               \
 416 })
 417
 418 static inline unsigned bset_u64s(struct bset_tree *t)
 419 {
 420         return t->end_offset - t->data_offset -
 421                 sizeof(struct bset) / sizeof(u64);
 422 }
 423
 424 static inline unsigned bset_dead_u64s(struct btree *b, struct bset_tree *t)
 425 {
 426         return bset_u64s(t) - b->nr.bset_u64s[t - b->set];
 427 }
 428
 429 static inline unsigned bset_byte_offset(struct btree *b, void *i)
 430 {
 431         return i - (void *) b->data;
 432 }
 433
 434 enum btree_node_type {
 435 #define x(kwd, val, name) BKEY_TYPE_##kwd = val,
 436         BCH_BTREE_IDS()
 437 #undef x
 438         BKEY_TYPE_BTREE,
 439 };
 440
 441 /* Type of a key in btree @id at level @level: */
 442 static inline enum btree_node_type __btree_node_type(unsigned level, enum btree_id id)
 443 {
 444         return level ? BKEY_TYPE_BTREE : (enum btree_node_type) id;
 445 }
 446
 447 /* Type of keys @b contains: */
 448 static inline enum btree_node_type btree_node_type(struct btree *b)
 449 {
 450         return __btree_node_type(b->level, b->btree_id);
 451 }
 452
 453 static inline bool btree_node_type_is_extents(enum btree_node_type type)
 454 {
 455         switch (type) {
 456         case BKEY_TYPE_EXTENTS:
 457         case BKEY_TYPE_REFLINK:
 458                 return true;
 459         default:
 460                 return false;
 461         }
 462 }
 463
 464 static inline bool btree_node_is_extents(struct btree *b)
 465 {
 466         return btree_node_type_is_extents(btree_node_type(b));
 467 }
 468
 469 #define BTREE_NODE_TYPE_HAS_TRIGGERS                    \
 470         ((1U << BKEY_TYPE_EXTENTS)|                     \
 471          (1U << BKEY_TYPE_ALLOC)|                       \
 472          (1U << BKEY_TYPE_INODES)|                      \
 473          (1U << BKEY_TYPE_REFLINK)|                     \
 474          (1U << BKEY_TYPE_EC)|                          \
 475          (1U << BKEY_TYPE_BTREE))
 476
 477 #define BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS              \
 478         ((1U << BKEY_TYPE_EXTENTS)|                     \
 479          (1U << BKEY_TYPE_INODES)|                      \
 480          (1U << BKEY_TYPE_REFLINK))
 481
 482 static inline bool btree_node_type_needs_gc(enum btree_node_type type)
 483 {
 484         return BTREE_NODE_TYPE_HAS_TRIGGERS & (1U << type);
 485 }
 486
 487 struct btree_root {
 488         struct btree            *b;
 489
 490         struct btree_update     *as;
 491
 492         /* On disk root - see async splits: */
 493         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 494         u8                      level;
 495         u8                      alive;
 496         s8                      error;
 497 };
 498
 499 /*
 500  * Optional hook that will be called just prior to a btree node update, when
 501  * we're holding the write lock and we know what key is about to be overwritten:
 502  */
 503
 504 enum btree_insert_ret {
 505         BTREE_INSERT_OK,
 506         /* leaf node needs to be split */
 507         BTREE_INSERT_BTREE_NODE_FULL,
 508         BTREE_INSERT_ENOSPC,
 509         BTREE_INSERT_NEED_MARK_REPLICAS,
 510         BTREE_INSERT_NEED_JOURNAL_RES,
 511 };
 512
 513 enum btree_gc_coalesce_fail_reason {
 514         BTREE_GC_COALESCE_FAIL_RESERVE_GET,
 515         BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC,
 516         BTREE_GC_COALESCE_FAIL_FORMAT_FITS,
 517 };
 518
 519 enum btree_node_sibling {
 520         btree_prev_sib,
 521         btree_next_sib,
 522 };
 523
 524 typedef struct btree_nr_keys (*sort_fix_overlapping_fn)(struct bset *,
 525                                                         struct btree *,
 526                                                         struct btree_node_iter *);
 527
 528 #endif /* _BCACHEFS_BTREE_TYPES_H */