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.
40 struct list_head list;
41 struct list_head unwritten_list;
43 /* What kind of update are we doing? */
45 BTREE_INTERIOR_NO_UPDATE,
46 BTREE_INTERIOR_UPDATING_NODE,
47 BTREE_INTERIOR_UPDATING_ROOT,
48 BTREE_INTERIOR_UPDATING_AS,
51 unsigned nodes_written:1;
52 unsigned took_gc_lock:1;
54 enum btree_id btree_id;
55 unsigned update_level;
57 struct disk_reservation disk_res;
60 * BTREE_INTERIOR_UPDATING_NODE:
61 * The update that made the new nodes visible was a regular update to an
62 * existing interior node - @b. We can't write out the update to @b
63 * until the new nodes we created are finished writing, so we block @b
64 * from writing by putting this btree_interior update on the
65 * @b->write_blocked list with @write_blocked_list:
68 struct list_head write_blocked_list;
71 * We may be freeing nodes that were dirty, and thus had journal entries
72 * pinned: we need to transfer the oldest of those pins to the
73 * btree_update operation, and release it when the new node(s)
74 * are all persistent and reachable:
76 struct journal_entry_pin journal;
78 /* Preallocated nodes we reserve when we start the update: */
79 struct prealloc_nodes {
80 struct btree *b[BTREE_UPDATE_NODES_MAX];
84 /* Nodes being freed: */
85 struct keylist old_keys;
86 u64 _old_keys[BTREE_UPDATE_NODES_MAX *
87 BKEY_BTREE_PTR_U64s_MAX];
89 /* Nodes being added: */
90 struct keylist new_keys;
91 u64 _new_keys[BTREE_UPDATE_NODES_MAX *
92 BKEY_BTREE_PTR_U64s_MAX];
94 /* New nodes, that will be made reachable by this update: */
95 struct btree *new_nodes[BTREE_UPDATE_NODES_MAX];
96 unsigned nr_new_nodes;
98 struct btree *old_nodes[BTREE_UPDATE_NODES_MAX];
99 __le64 old_nodes_seq[BTREE_UPDATE_NODES_MAX];
100 unsigned nr_old_nodes;
102 open_bucket_idx_t open_buckets[BTREE_UPDATE_NODES_MAX *
104 open_bucket_idx_t nr_open_buckets;
106 unsigned journal_u64s;
107 u64 journal_entries[BTREE_UPDATE_JOURNAL_RES];
109 /* Only here to reduce stack usage on recursive splits: */
110 struct keylist parent_keys;
112 * Enough room for btree_split's keys without realloc - btree node
113 * pointers never have crc/compression info, so we only need to acount
114 * for the pointers for three keys
116 u64 inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
119 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
120 struct btree_trans *,
124 int bch2_btree_split_leaf(struct btree_trans *, struct btree_path *, unsigned);
126 int __bch2_foreground_maybe_merge(struct btree_trans *, struct btree_path *,
127 unsigned, unsigned, enum btree_node_sibling);
129 static inline int bch2_foreground_maybe_merge_sibling(struct btree_trans *trans,
130 struct btree_path *path,
131 unsigned level, unsigned flags,
132 enum btree_node_sibling sib)
136 EBUG_ON(!btree_node_locked(path, level));
138 b = path->l[level].b;
139 if (b->sib_u64s[sib] > trans->c->btree_foreground_merge_threshold)
142 return __bch2_foreground_maybe_merge(trans, path, level, flags, sib);
145 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
146 struct btree_path *path,
150 return bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
152 bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
156 int bch2_btree_node_rewrite(struct btree_trans *, struct btree_iter *,
157 struct btree *, unsigned);
158 void bch2_btree_node_rewrite_async(struct bch_fs *, struct btree *);
159 int bch2_btree_node_update_key(struct btree_trans *, struct btree_iter *,
160 struct btree *, struct bkey_i *,
162 int bch2_btree_node_update_key_get_iter(struct btree_trans *, struct btree *,
163 struct bkey_i *, unsigned, bool);
165 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
166 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
168 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
171 unsigned depth = btree_node_root(c, b)->c.level + 1;
174 * Number of nodes we might have to allocate in a worst case btree
175 * split operation - we split all the way up to the root, then allocate
176 * a new root, unless we're already at max depth:
178 if (depth < BTREE_MAX_DEPTH)
179 return (depth - b->c.level) * 2 + 1;
181 return (depth - b->c.level) * 2 - 1;
184 static inline void btree_node_reset_sib_u64s(struct btree *b)
186 b->sib_u64s[0] = b->nr.live_u64s;
187 b->sib_u64s[1] = b->nr.live_u64s;
190 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
192 return (void *) b->data + btree_bytes(c);
195 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
198 return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
201 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
204 return btree_data_end(c, b);
207 static inline void *write_block(struct btree *b)
209 return (void *) b->data + (b->written << 9);
212 static inline bool __btree_addr_written(struct btree *b, void *p)
214 return p < write_block(b);
217 static inline bool bset_written(struct btree *b, struct bset *i)
219 return __btree_addr_written(b, i);
222 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
224 return __btree_addr_written(b, k);
227 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
231 ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
233 ssize_t total = c->opts.btree_node_size >> 3;
235 /* Always leave one extra u64 for bch2_varint_decode: */
241 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
244 ssize_t remaining = __bch_btree_u64s_remaining(c, b,
245 btree_bkey_last(b, bset_tree_last(b)));
247 BUG_ON(remaining < 0);
249 if (bset_written(b, btree_bset_last(b)))
255 #define BTREE_WRITE_SET_U64s_BITS 9
257 static inline unsigned btree_write_set_buffer(struct btree *b)
260 * Could buffer up larger amounts of keys for btrees with larger keys,
261 * pending benchmarking:
263 return 8 << BTREE_WRITE_SET_U64s_BITS;
266 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
269 struct bset_tree *t = bset_tree_last(b);
270 struct btree_node_entry *bne = max(write_block(b),
271 (void *) btree_bkey_last(b, bset_tree_last(b)));
272 ssize_t remaining_space =
273 __bch_btree_u64s_remaining(c, b, bne->keys.start);
275 if (unlikely(bset_written(b, bset(b, t)))) {
276 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
279 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
280 remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
287 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
290 struct bkey_packed k;
292 BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
293 EBUG_ON(btree_node_just_written(b));
295 if (!bkey_pack_pos(&k, pos, b)) {
296 struct bkey *u = (void *) &k;
302 k.needs_whiteout = true;
304 b->whiteout_u64s += k.u64s;
305 bkey_p_copy(unwritten_whiteouts_start(c, b), &k);
309 * write lock must be held on @b (else the dirty bset that we were going to
310 * insert into could be written out from under us)
312 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
313 struct btree *b, unsigned u64s)
315 if (unlikely(btree_node_need_rewrite(b)))
318 return u64s <= bch_btree_keys_u64s_remaining(c, b);
321 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
323 bool bch2_btree_interior_updates_flush(struct bch_fs *);
325 void bch2_journal_entry_to_btree_root(struct bch_fs *, struct jset_entry *);
326 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
327 struct jset_entry *, unsigned long);
329 void bch2_do_pending_node_rewrites(struct bch_fs *);
330 void bch2_free_pending_node_rewrites(struct bch_fs *);
332 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
333 void bch2_fs_btree_interior_update_init_early(struct bch_fs *);
334 int bch2_fs_btree_interior_update_init(struct bch_fs *);
336 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */