347 lines · c
1/* SPDX-License-Identifier: GPL-2.0 */2#ifndef _BCACHEFS_BTREE_UPDATE_INTERIOR_H3#define _BCACHEFS_BTREE_UPDATE_INTERIOR_H4 5#include "btree_cache.h"6#include "btree_locking.h"7#include "btree_update.h"8 9#define BTREE_UPDATE_NODES_MAX ((BTREE_MAX_DEPTH - 2) * 2 + GC_MERGE_NODES)10 11#define BTREE_UPDATE_JOURNAL_RES (BTREE_UPDATE_NODES_MAX * (BKEY_BTREE_PTR_U64s_MAX + 1))12 13int bch2_btree_node_check_topology(struct btree_trans *, struct btree *);14 15#define BTREE_UPDATE_MODES() \16 x(none) \17 x(node) \18 x(root) \19 x(update)20 21enum btree_update_mode {22#define x(n) BTREE_UPDATE_##n,23 BTREE_UPDATE_MODES()24#undef x25};26 27/*28 * Tracks an in progress split/rewrite of a btree node and the update to the29 * parent node:30 *31 * When we split/rewrite a node, we do all the updates in memory without32 * waiting for any writes to complete - we allocate the new node(s) and update33 * the parent node, possibly recursively up to the root.34 *35 * The end result is that we have one or more new nodes being written -36 * possibly several, if there were multiple splits - and then a write (updating37 * an interior node) which will make all these new nodes visible.38 *39 * Additionally, as we split/rewrite nodes we free the old nodes - but the old40 * nodes can't be freed (their space on disk can't be reclaimed) until the41 * update to the interior node that makes the new node visible completes -42 * until then, the old nodes are still reachable on disk.43 *44 */45struct btree_update {46 struct closure cl;47 struct bch_fs *c;48 u64 start_time;49 unsigned long ip_started;50 51 struct list_head list;52 struct list_head unwritten_list;53 54 enum btree_update_mode mode;55 enum bch_trans_commit_flags flags;56 unsigned nodes_written:1;57 unsigned took_gc_lock:1;58 59 enum btree_id btree_id;60 unsigned update_level_start;61 unsigned update_level_end;62 63 struct disk_reservation disk_res;64 65 /*66 * BTREE_UPDATE_node:67 * The update that made the new nodes visible was a regular update to an68 * existing interior node - @b. We can't write out the update to @b69 * until the new nodes we created are finished writing, so we block @b70 * from writing by putting this btree_interior update on the71 * @b->write_blocked list with @write_blocked_list:72 */73 struct btree *b;74 struct list_head write_blocked_list;75 76 /*77 * We may be freeing nodes that were dirty, and thus had journal entries78 * pinned: we need to transfer the oldest of those pins to the79 * btree_update operation, and release it when the new node(s)80 * are all persistent and reachable:81 */82 struct journal_entry_pin journal;83 84 /* Preallocated nodes we reserve when we start the update: */85 struct prealloc_nodes {86 struct btree *b[BTREE_UPDATE_NODES_MAX];87 unsigned nr;88 } prealloc_nodes[2];89 90 /* Nodes being freed: */91 struct keylist old_keys;92 u64 _old_keys[BTREE_UPDATE_NODES_MAX *93 BKEY_BTREE_PTR_U64s_MAX];94 95 /* Nodes being added: */96 struct keylist new_keys;97 u64 _new_keys[BTREE_UPDATE_NODES_MAX *98 BKEY_BTREE_PTR_U64s_MAX];99 100 /* New nodes, that will be made reachable by this update: */101 struct btree *new_nodes[BTREE_UPDATE_NODES_MAX];102 unsigned nr_new_nodes;103 104 struct btree *old_nodes[BTREE_UPDATE_NODES_MAX];105 __le64 old_nodes_seq[BTREE_UPDATE_NODES_MAX];106 unsigned nr_old_nodes;107 108 open_bucket_idx_t open_buckets[BTREE_UPDATE_NODES_MAX *109 BCH_REPLICAS_MAX];110 open_bucket_idx_t nr_open_buckets;111 112 unsigned journal_u64s;113 u64 journal_entries[BTREE_UPDATE_JOURNAL_RES];114 115 /* Only here to reduce stack usage on recursive splits: */116 struct keylist parent_keys;117 /*118 * Enough room for btree_split's keys without realloc - btree node119 * pointers never have crc/compression info, so we only need to acount120 * for the pointers for three keys121 */122 u64 inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];123};124 125struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,126 struct btree_trans *,127 struct btree *,128 struct bkey_format);129 130int bch2_btree_split_leaf(struct btree_trans *, btree_path_idx_t, unsigned);131 132int bch2_btree_increase_depth(struct btree_trans *, btree_path_idx_t, unsigned);133 134int __bch2_foreground_maybe_merge(struct btree_trans *, btree_path_idx_t,135 unsigned, unsigned, enum btree_node_sibling);136 137static inline int bch2_foreground_maybe_merge_sibling(struct btree_trans *trans,138 btree_path_idx_t path_idx,139 unsigned level, unsigned flags,140 enum btree_node_sibling sib)141{142 struct btree_path *path = trans->paths + path_idx;143 struct btree *b;144 145 EBUG_ON(!btree_node_locked(path, level));146 147 if (bch2_btree_node_merging_disabled)148 return 0;149 150 b = path->l[level].b;151 if (b->sib_u64s[sib] > trans->c->btree_foreground_merge_threshold)152 return 0;153 154 return __bch2_foreground_maybe_merge(trans, path_idx, level, flags, sib);155}156 157static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,158 btree_path_idx_t path,159 unsigned level,160 unsigned flags)161{162 bch2_trans_verify_not_unlocked(trans);163 164 return bch2_foreground_maybe_merge_sibling(trans, path, level, flags,165 btree_prev_sib) ?:166 bch2_foreground_maybe_merge_sibling(trans, path, level, flags,167 btree_next_sib);168}169 170int bch2_btree_node_rewrite(struct btree_trans *, struct btree_iter *,171 struct btree *, unsigned);172void bch2_btree_node_rewrite_async(struct bch_fs *, struct btree *);173int bch2_btree_node_update_key(struct btree_trans *, struct btree_iter *,174 struct btree *, struct bkey_i *,175 unsigned, bool);176int bch2_btree_node_update_key_get_iter(struct btree_trans *, struct btree *,177 struct bkey_i *, unsigned, bool);178 179void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);180 181int bch2_btree_root_alloc_fake_trans(struct btree_trans *, enum btree_id, unsigned);182void bch2_btree_root_alloc_fake(struct bch_fs *, enum btree_id, unsigned);183 184static inline unsigned btree_update_reserve_required(struct bch_fs *c,185 struct btree *b)186{187 unsigned depth = btree_node_root(c, b)->c.level + 1;188 189 /*190 * Number of nodes we might have to allocate in a worst case btree191 * split operation - we split all the way up to the root, then allocate192 * a new root, unless we're already at max depth:193 */194 if (depth < BTREE_MAX_DEPTH)195 return (depth - b->c.level) * 2 + 1;196 else197 return (depth - b->c.level) * 2 - 1;198}199 200static inline void btree_node_reset_sib_u64s(struct btree *b)201{202 b->sib_u64s[0] = b->nr.live_u64s;203 b->sib_u64s[1] = b->nr.live_u64s;204}205 206static inline void *btree_data_end(struct btree *b)207{208 return (void *) b->data + btree_buf_bytes(b);209}210 211static inline struct bkey_packed *unwritten_whiteouts_start(struct btree *b)212{213 return (void *) ((u64 *) btree_data_end(b) - b->whiteout_u64s);214}215 216static inline struct bkey_packed *unwritten_whiteouts_end(struct btree *b)217{218 return btree_data_end(b);219}220 221static inline void *write_block(struct btree *b)222{223 return (void *) b->data + (b->written << 9);224}225 226static inline bool __btree_addr_written(struct btree *b, void *p)227{228 return p < write_block(b);229}230 231static inline bool bset_written(struct btree *b, struct bset *i)232{233 return __btree_addr_written(b, i);234}235 236static inline bool bkey_written(struct btree *b, struct bkey_packed *k)237{238 return __btree_addr_written(b, k);239}240 241static inline ssize_t __bch2_btree_u64s_remaining(struct btree *b, void *end)242{243 ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +244 b->whiteout_u64s;245 ssize_t total = btree_buf_bytes(b) >> 3;246 247 /* Always leave one extra u64 for bch2_varint_decode: */248 used++;249 250 return total - used;251}252 253static inline size_t bch2_btree_keys_u64s_remaining(struct btree *b)254{255 ssize_t remaining = __bch2_btree_u64s_remaining(b,256 btree_bkey_last(b, bset_tree_last(b)));257 258 BUG_ON(remaining < 0);259 260 if (bset_written(b, btree_bset_last(b)))261 return 0;262 263 return remaining;264}265 266#define BTREE_WRITE_SET_U64s_BITS 9267 268static inline unsigned btree_write_set_buffer(struct btree *b)269{270 /*271 * Could buffer up larger amounts of keys for btrees with larger keys,272 * pending benchmarking:273 */274 return 8 << BTREE_WRITE_SET_U64s_BITS;275}276 277static inline struct btree_node_entry *want_new_bset(struct bch_fs *c, struct btree *b)278{279 struct bset_tree *t = bset_tree_last(b);280 struct btree_node_entry *bne = max(write_block(b),281 (void *) btree_bkey_last(b, bset_tree_last(b)));282 ssize_t remaining_space =283 __bch2_btree_u64s_remaining(b, bne->keys.start);284 285 if (unlikely(bset_written(b, bset(b, t)))) {286 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))287 return bne;288 } else {289 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&290 remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))291 return bne;292 }293 294 return NULL;295}296 297static inline void push_whiteout(struct btree *b, struct bpos pos)298{299 struct bkey_packed k;300 301 BUG_ON(bch2_btree_keys_u64s_remaining(b) < BKEY_U64s);302 EBUG_ON(btree_node_just_written(b));303 304 if (!bkey_pack_pos(&k, pos, b)) {305 struct bkey *u = (void *) &k;306 307 bkey_init(u);308 u->p = pos;309 }310 311 k.needs_whiteout = true;312 313 b->whiteout_u64s += k.u64s;314 bkey_p_copy(unwritten_whiteouts_start(b), &k);315}316 317/*318 * write lock must be held on @b (else the dirty bset that we were going to319 * insert into could be written out from under us)320 */321static inline bool bch2_btree_node_insert_fits(struct btree *b, unsigned u64s)322{323 if (unlikely(btree_node_need_rewrite(b)))324 return false;325 326 return u64s <= bch2_btree_keys_u64s_remaining(b);327}328 329void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);330 331bool bch2_btree_interior_updates_flush(struct bch_fs *);332 333void bch2_journal_entry_to_btree_root(struct bch_fs *, struct jset_entry *);334struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,335 struct jset_entry *, unsigned long);336 337void bch2_do_pending_node_rewrites(struct bch_fs *);338void bch2_free_pending_node_rewrites(struct bch_fs *);339 340void bch2_btree_reserve_cache_to_text(struct printbuf *, struct bch_fs *);341 342void bch2_fs_btree_interior_update_exit(struct bch_fs *);343void bch2_fs_btree_interior_update_init_early(struct bch_fs *);344int bch2_fs_btree_interior_update_init(struct bch_fs *);345 346#endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */347