1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _BCACHEFS_BTREE_UPDATE_INTERIOR_H
3 #define _BCACHEFS_BTREE_UPDATE_INTERIOR_H
5 #include "btree_cache.h"
6 #include "btree_locking.h"
7 #include "btree_update.h"
9 void __bch2_btree_calc_format(struct bkey_format_state *, struct btree *);
10 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *,
11 struct bkey_format *);
13 #define BTREE_UPDATE_NODES_MAX ((BTREE_MAX_DEPTH - 2) * 2 + GC_MERGE_NODES)
15 #define BTREE_UPDATE_JOURNAL_RES (BTREE_UPDATE_NODES_MAX * (BKEY_BTREE_PTR_U64s_MAX + 1))
18 * Tracks an in progress split/rewrite of a btree node and the update to the
21 * When we split/rewrite a node, we do all the updates in memory without
22 * waiting for any writes to complete - we allocate the new node(s) and update
23 * the parent node, possibly recursively up to the root.
25 * The end result is that we have one or more new nodes being written -
26 * possibly several, if there were multiple splits - and then a write (updating
27 * an interior node) which will make all these new nodes visible.
29 * Additionally, as we split/rewrite nodes we free the old nodes - but the old
30 * nodes can't be freed (their space on disk can't be reclaimed) until the
31 * update to the interior node that makes the new node visible completes -
32 * until then, the old nodes are still reachable on disk.
39 struct list_head list;
40 struct list_head unwritten_list;
42 /* What kind of update are we doing? */
44 BTREE_INTERIOR_NO_UPDATE,
45 BTREE_INTERIOR_UPDATING_NODE,
46 BTREE_INTERIOR_UPDATING_ROOT,
47 BTREE_INTERIOR_UPDATING_AS,
50 unsigned nodes_written:1;
52 enum btree_id btree_id;
54 struct disk_reservation disk_res;
55 struct journal_preres journal_preres;
58 * BTREE_INTERIOR_UPDATING_NODE:
59 * The update that made the new nodes visible was a regular update to an
60 * existing interior node - @b. We can't write out the update to @b
61 * until the new nodes we created are finished writing, so we block @b
62 * from writing by putting this btree_interior update on the
63 * @b->write_blocked list with @write_blocked_list:
66 struct list_head write_blocked_list;
69 * We may be freeing nodes that were dirty, and thus had journal entries
70 * pinned: we need to transfer the oldest of those pins to the
71 * btree_update operation, and release it when the new node(s)
72 * are all persistent and reachable:
74 struct journal_entry_pin journal;
76 /* Preallocated nodes we reserve when we start the update: */
77 struct btree *prealloc_nodes[BTREE_UPDATE_NODES_MAX];
78 unsigned nr_prealloc_nodes;
80 /* Nodes being freed: */
81 struct keylist old_keys;
82 u64 _old_keys[BTREE_UPDATE_NODES_MAX *
83 BKEY_BTREE_PTR_VAL_U64s_MAX];
85 /* Nodes being added: */
86 struct keylist new_keys;
87 u64 _new_keys[BTREE_UPDATE_NODES_MAX *
88 BKEY_BTREE_PTR_VAL_U64s_MAX];
90 /* New nodes, that will be made reachable by this update: */
91 struct btree *new_nodes[BTREE_UPDATE_NODES_MAX];
92 unsigned nr_new_nodes;
94 open_bucket_idx_t open_buckets[BTREE_UPDATE_NODES_MAX *
96 open_bucket_idx_t nr_open_buckets;
98 unsigned journal_u64s;
99 u64 journal_entries[BTREE_UPDATE_JOURNAL_RES];
101 /* Only here to reduce stack usage on recursive splits: */
102 struct keylist parent_keys;
104 * Enough room for btree_split's keys without realloc - btree node
105 * pointers never have crc/compression info, so we only need to acount
106 * for the pointers for three keys
108 u64 inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
111 void bch2_btree_node_free_inmem(struct bch_fs *, struct btree *,
112 struct btree_iter *);
113 void bch2_btree_node_free_never_inserted(struct bch_fs *, struct btree *);
115 void bch2_btree_update_get_open_buckets(struct btree_update *, struct btree *);
117 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
121 void bch2_btree_update_done(struct btree_update *);
122 struct btree_update *
123 bch2_btree_update_start(struct btree_trans *, enum btree_id, unsigned,
124 unsigned, struct closure *);
126 void bch2_btree_interior_update_will_free_node(struct btree_update *,
128 void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
130 void bch2_btree_insert_node(struct btree_update *, struct btree *,
131 struct btree_iter *, struct keylist *,
133 int bch2_btree_split_leaf(struct bch_fs *, struct btree_iter *, unsigned);
135 void __bch2_foreground_maybe_merge(struct bch_fs *, struct btree_iter *,
136 unsigned, unsigned, enum btree_node_sibling);
138 static inline void bch2_foreground_maybe_merge_sibling(struct bch_fs *c,
139 struct btree_iter *iter,
140 unsigned level, unsigned flags,
141 enum btree_node_sibling sib)
145 if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
148 if (!bch2_btree_node_relock(iter, level))
151 b = iter->l[level].b;
152 if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
155 __bch2_foreground_maybe_merge(c, iter, level, flags, sib);
158 static inline void bch2_foreground_maybe_merge(struct bch_fs *c,
159 struct btree_iter *iter,
163 bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
165 bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
169 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
170 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
172 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
175 unsigned depth = btree_node_root(c, b)->c.level + 1;
178 * Number of nodes we might have to allocate in a worst case btree
179 * split operation - we split all the way up to the root, then allocate
180 * a new root, unless we're already at max depth:
182 if (depth < BTREE_MAX_DEPTH)
183 return (depth - b->c.level) * 2 + 1;
185 return (depth - b->c.level) * 2 - 1;
188 static inline void btree_node_reset_sib_u64s(struct btree *b)
190 b->sib_u64s[0] = b->nr.live_u64s;
191 b->sib_u64s[1] = b->nr.live_u64s;
194 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
196 return (void *) b->data + btree_bytes(c);
199 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
202 return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
205 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
208 return btree_data_end(c, b);
211 static inline void *write_block(struct btree *b)
213 return (void *) b->data + (b->written << 9);
216 static inline bool __btree_addr_written(struct btree *b, void *p)
218 return p < write_block(b);
221 static inline bool bset_written(struct btree *b, struct bset *i)
223 return __btree_addr_written(b, i);
226 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
228 return __btree_addr_written(b, k);
231 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
235 ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
237 ssize_t total = c->opts.btree_node_size << 6;
239 /* Always leave one extra u64 for bch2_varint_decode: */
245 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
248 ssize_t remaining = __bch_btree_u64s_remaining(c, b,
249 btree_bkey_last(b, bset_tree_last(b)));
251 BUG_ON(remaining < 0);
253 if (bset_written(b, btree_bset_last(b)))
259 static inline unsigned btree_write_set_buffer(struct btree *b)
262 * Could buffer up larger amounts of keys for btrees with larger keys,
263 * pending benchmarking:
268 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
271 struct bset_tree *t = bset_tree_last(b);
272 struct btree_node_entry *bne = max(write_block(b),
273 (void *) btree_bkey_last(b, bset_tree_last(b)));
274 ssize_t remaining_space =
275 __bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
277 if (unlikely(bset_written(b, bset(b, t)))) {
278 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
281 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
282 remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
289 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
292 struct bkey_packed k;
294 BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
296 if (!bkey_pack_pos(&k, pos, b)) {
297 struct bkey *u = (void *) &k;
303 k.needs_whiteout = true;
305 b->whiteout_u64s += k.u64s;
306 bkey_copy(unwritten_whiteouts_start(c, b), &k);
310 * write lock must be held on @b (else the dirty bset that we were going to
311 * insert into could be written out from under us)
313 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
314 struct btree *b, unsigned u64s)
316 if (unlikely(btree_node_need_rewrite(b)))
319 return u64s <= bch_btree_keys_u64s_remaining(c, b);
322 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
324 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *);
326 void bch2_journal_entries_to_btree_roots(struct bch_fs *, struct jset *);
327 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
328 struct jset_entry *, struct jset_entry *);
330 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
331 int bch2_fs_btree_interior_update_init(struct bch_fs *);
333 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */