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;
58 struct journal_preres journal_preres;
61 * BTREE_INTERIOR_UPDATING_NODE:
62 * The update that made the new nodes visible was a regular update to an
63 * existing interior node - @b. We can't write out the update to @b
64 * until the new nodes we created are finished writing, so we block @b
65 * from writing by putting this btree_interior update on the
66 * @b->write_blocked list with @write_blocked_list:
69 struct list_head write_blocked_list;
72 * We may be freeing nodes that were dirty, and thus had journal entries
73 * pinned: we need to transfer the oldest of those pins to the
74 * btree_update operation, and release it when the new node(s)
75 * are all persistent and reachable:
77 struct journal_entry_pin journal;
79 /* Preallocated nodes we reserve when we start the update: */
80 struct prealloc_nodes {
81 struct btree *b[BTREE_UPDATE_NODES_MAX];
85 /* Nodes being freed: */
86 struct keylist old_keys;
87 u64 _old_keys[BTREE_UPDATE_NODES_MAX *
88 BKEY_BTREE_PTR_U64s_MAX];
90 /* Nodes being added: */
91 struct keylist new_keys;
92 u64 _new_keys[BTREE_UPDATE_NODES_MAX *
93 BKEY_BTREE_PTR_U64s_MAX];
95 /* New nodes, that will be made reachable by this update: */
96 struct btree *new_nodes[BTREE_UPDATE_NODES_MAX];
97 unsigned nr_new_nodes;
99 struct btree *old_nodes[BTREE_UPDATE_NODES_MAX];
100 __le64 old_nodes_seq[BTREE_UPDATE_NODES_MAX];
101 unsigned nr_old_nodes;
103 open_bucket_idx_t open_buckets[BTREE_UPDATE_NODES_MAX *
105 open_bucket_idx_t nr_open_buckets;
107 unsigned journal_u64s;
108 u64 journal_entries[BTREE_UPDATE_JOURNAL_RES];
110 /* Only here to reduce stack usage on recursive splits: */
111 struct keylist parent_keys;
113 * Enough room for btree_split's keys without realloc - btree node
114 * pointers never have crc/compression info, so we only need to acount
115 * for the pointers for three keys
117 u64 inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
120 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
121 struct btree_trans *,
125 int bch2_btree_split_leaf(struct btree_trans *, struct btree_path *, unsigned);
127 int __bch2_foreground_maybe_merge(struct btree_trans *, struct btree_path *,
128 unsigned, unsigned, enum btree_node_sibling);
130 static inline int bch2_foreground_maybe_merge_sibling(struct btree_trans *trans,
131 struct btree_path *path,
132 unsigned level, unsigned flags,
133 enum btree_node_sibling sib)
137 EBUG_ON(!btree_node_locked(path, level));
139 b = path->l[level].b;
140 if (b->sib_u64s[sib] > trans->c->btree_foreground_merge_threshold)
143 return __bch2_foreground_maybe_merge(trans, path, level, flags, sib);
146 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
147 struct btree_path *path,
151 return bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
153 bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
157 int bch2_btree_node_rewrite(struct btree_trans *, struct btree_iter *,
158 struct btree *, unsigned);
159 void bch2_btree_node_rewrite_async(struct bch_fs *, struct btree *);
160 int bch2_btree_node_update_key(struct btree_trans *, struct btree_iter *,
161 struct btree *, struct bkey_i *,
163 int bch2_btree_node_update_key_get_iter(struct btree_trans *, struct btree *,
164 struct bkey_i *, unsigned, bool);
166 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
167 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
169 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
172 unsigned depth = btree_node_root(c, b)->c.level + 1;
175 * Number of nodes we might have to allocate in a worst case btree
176 * split operation - we split all the way up to the root, then allocate
177 * a new root, unless we're already at max depth:
179 if (depth < BTREE_MAX_DEPTH)
180 return (depth - b->c.level) * 2 + 1;
182 return (depth - b->c.level) * 2 - 1;
185 static inline void btree_node_reset_sib_u64s(struct btree *b)
187 b->sib_u64s[0] = b->nr.live_u64s;
188 b->sib_u64s[1] = b->nr.live_u64s;
191 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
193 return (void *) b->data + btree_bytes(c);
196 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
199 return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
202 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
205 return btree_data_end(c, b);
208 static inline void *write_block(struct btree *b)
210 return (void *) b->data + (b->written << 9);
213 static inline bool __btree_addr_written(struct btree *b, void *p)
215 return p < write_block(b);
218 static inline bool bset_written(struct btree *b, struct bset *i)
220 return __btree_addr_written(b, i);
223 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
225 return __btree_addr_written(b, k);
228 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
232 ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
234 ssize_t total = c->opts.btree_node_size >> 3;
236 /* Always leave one extra u64 for bch2_varint_decode: */
242 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
245 ssize_t remaining = __bch_btree_u64s_remaining(c, b,
246 btree_bkey_last(b, bset_tree_last(b)));
248 BUG_ON(remaining < 0);
250 if (bset_written(b, btree_bset_last(b)))
256 #define BTREE_WRITE_SET_U64s_BITS 9
258 static inline unsigned btree_write_set_buffer(struct btree *b)
261 * Could buffer up larger amounts of keys for btrees with larger keys,
262 * pending benchmarking:
264 return 8 << BTREE_WRITE_SET_U64s_BITS;
267 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
270 struct bset_tree *t = bset_tree_last(b);
271 struct btree_node_entry *bne = max(write_block(b),
272 (void *) btree_bkey_last(b, bset_tree_last(b)));
273 ssize_t remaining_space =
274 __bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
276 if (unlikely(bset_written(b, bset(b, t)))) {
277 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
280 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
281 remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
288 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
291 struct bkey_packed k;
293 BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
294 EBUG_ON(btree_node_just_written(b));
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 bool bch2_btree_interior_updates_flush(struct bch_fs *);
326 void bch2_journal_entry_to_btree_root(struct bch_fs *, struct jset_entry *);
327 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
328 struct jset_entry *, struct jset_entry *);
330 void bch2_do_pending_node_rewrites(struct bch_fs *);
331 void bch2_free_pending_node_rewrites(struct bch_fs *);
333 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
334 void bch2_fs_btree_interior_update_init_early(struct bch_fs *);
335 int bch2_fs_btree_interior_update_init(struct bch_fs *);
337 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */