xref: /linux/fs/bcachefs/btree_update_interior.h (revision 92815da4576a495cb6362cdfb132152fccc2222d)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _BCACHEFS_BTREE_UPDATE_INTERIOR_H
3 #define _BCACHEFS_BTREE_UPDATE_INTERIOR_H
4 
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 
13 int 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 
21 enum btree_update_mode {
22 #define x(n)	BTREE_UPDATE_##n,
23 	BTREE_UPDATE_MODES()
24 #undef x
25 };
26 
27 /*
28  * Tracks an in progress split/rewrite of a btree node and the update to the
29  * parent node:
30  *
31  * When we split/rewrite a node, we do all the updates in memory without
32  * waiting for any writes to complete - we allocate the new node(s) and update
33  * 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 (updating
37  * 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 old
40  * nodes can't be freed (their space on disk can't be reclaimed) until the
41  * 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  */
45 struct 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 an
68 	 * existing interior node - @b. We can't write out the update to @b
69 	 * until the new nodes we created are finished writing, so we block @b
70 	 * from writing by putting this btree_interior update on the
71 	 * @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 entries
78 	 * pinned: we need to transfer the oldest of those pins to the
79 	 * 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 node
119 	 * pointers never have crc/compression info, so we only need to acount
120 	 * for the pointers for three keys
121 	 */
122 	u64				inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
123 };
124 
125 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
126 						  struct btree_trans *,
127 						  struct btree *,
128 						  struct bkey_format);
129 
130 int bch2_btree_split_leaf(struct btree_trans *, btree_path_idx_t, unsigned);
131 
132 int bch2_btree_increase_depth(struct btree_trans *, btree_path_idx_t, unsigned);
133 
134 int __bch2_foreground_maybe_merge(struct btree_trans *, btree_path_idx_t,
135 				  unsigned, unsigned, enum btree_node_sibling);
136 
137 static 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 
157 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
158 					      btree_path_idx_t path,
159 					      unsigned level,
160 					      unsigned flags)
161 {
162 	return  bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
163 						    btree_prev_sib) ?:
164 		bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
165 						    btree_next_sib);
166 }
167 
168 int bch2_btree_node_rewrite(struct btree_trans *, struct btree_iter *,
169 			    struct btree *, unsigned);
170 void bch2_btree_node_rewrite_async(struct bch_fs *, struct btree *);
171 int bch2_btree_node_update_key(struct btree_trans *, struct btree_iter *,
172 			       struct btree *, struct bkey_i *,
173 			       unsigned, bool);
174 int bch2_btree_node_update_key_get_iter(struct btree_trans *, struct btree *,
175 					struct bkey_i *, unsigned, bool);
176 
177 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
178 
179 int bch2_btree_root_alloc_fake_trans(struct btree_trans *, enum btree_id, unsigned);
180 void bch2_btree_root_alloc_fake(struct bch_fs *, enum btree_id, unsigned);
181 
182 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
183 						     struct btree *b)
184 {
185 	unsigned depth = btree_node_root(c, b)->c.level + 1;
186 
187 	/*
188 	 * Number of nodes we might have to allocate in a worst case btree
189 	 * split operation - we split all the way up to the root, then allocate
190 	 * a new root, unless we're already at max depth:
191 	 */
192 	if (depth < BTREE_MAX_DEPTH)
193 		return (depth - b->c.level) * 2 + 1;
194 	else
195 		return (depth - b->c.level) * 2 - 1;
196 }
197 
198 static inline void btree_node_reset_sib_u64s(struct btree *b)
199 {
200 	b->sib_u64s[0] = b->nr.live_u64s;
201 	b->sib_u64s[1] = b->nr.live_u64s;
202 }
203 
204 static inline void *btree_data_end(struct btree *b)
205 {
206 	return (void *) b->data + btree_buf_bytes(b);
207 }
208 
209 static inline struct bkey_packed *unwritten_whiteouts_start(struct btree *b)
210 {
211 	return (void *) ((u64 *) btree_data_end(b) - b->whiteout_u64s);
212 }
213 
214 static inline struct bkey_packed *unwritten_whiteouts_end(struct btree *b)
215 {
216 	return btree_data_end(b);
217 }
218 
219 static inline void *write_block(struct btree *b)
220 {
221 	return (void *) b->data + (b->written << 9);
222 }
223 
224 static inline bool __btree_addr_written(struct btree *b, void *p)
225 {
226 	return p < write_block(b);
227 }
228 
229 static inline bool bset_written(struct btree *b, struct bset *i)
230 {
231 	return __btree_addr_written(b, i);
232 }
233 
234 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
235 {
236 	return __btree_addr_written(b, k);
237 }
238 
239 static inline ssize_t __bch2_btree_u64s_remaining(struct btree *b, void *end)
240 {
241 	ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
242 		b->whiteout_u64s;
243 	ssize_t total = btree_buf_bytes(b) >> 3;
244 
245 	/* Always leave one extra u64 for bch2_varint_decode: */
246 	used++;
247 
248 	return total - used;
249 }
250 
251 static inline size_t bch2_btree_keys_u64s_remaining(struct btree *b)
252 {
253 	ssize_t remaining = __bch2_btree_u64s_remaining(b,
254 				btree_bkey_last(b, bset_tree_last(b)));
255 
256 	BUG_ON(remaining < 0);
257 
258 	if (bset_written(b, btree_bset_last(b)))
259 		return 0;
260 
261 	return remaining;
262 }
263 
264 #define BTREE_WRITE_SET_U64s_BITS	9
265 
266 static inline unsigned btree_write_set_buffer(struct btree *b)
267 {
268 	/*
269 	 * Could buffer up larger amounts of keys for btrees with larger keys,
270 	 * pending benchmarking:
271 	 */
272 	return 8 << BTREE_WRITE_SET_U64s_BITS;
273 }
274 
275 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c, struct btree *b)
276 {
277 	struct bset_tree *t = bset_tree_last(b);
278 	struct btree_node_entry *bne = max(write_block(b),
279 			(void *) btree_bkey_last(b, bset_tree_last(b)));
280 	ssize_t remaining_space =
281 		__bch2_btree_u64s_remaining(b, bne->keys.start);
282 
283 	if (unlikely(bset_written(b, bset(b, t)))) {
284 		if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
285 			return bne;
286 	} else {
287 		if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
288 		    remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
289 			return bne;
290 	}
291 
292 	return NULL;
293 }
294 
295 static inline void push_whiteout(struct btree *b, struct bpos pos)
296 {
297 	struct bkey_packed k;
298 
299 	BUG_ON(bch2_btree_keys_u64s_remaining(b) < BKEY_U64s);
300 	EBUG_ON(btree_node_just_written(b));
301 
302 	if (!bkey_pack_pos(&k, pos, b)) {
303 		struct bkey *u = (void *) &k;
304 
305 		bkey_init(u);
306 		u->p = pos;
307 	}
308 
309 	k.needs_whiteout = true;
310 
311 	b->whiteout_u64s += k.u64s;
312 	bkey_p_copy(unwritten_whiteouts_start(b), &k);
313 }
314 
315 /*
316  * write lock must be held on @b (else the dirty bset that we were going to
317  * insert into could be written out from under us)
318  */
319 static inline bool bch2_btree_node_insert_fits(struct btree *b, unsigned u64s)
320 {
321 	if (unlikely(btree_node_need_rewrite(b)))
322 		return false;
323 
324 	return u64s <= bch2_btree_keys_u64s_remaining(b);
325 }
326 
327 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
328 
329 bool bch2_btree_interior_updates_flush(struct bch_fs *);
330 
331 void bch2_journal_entry_to_btree_root(struct bch_fs *, struct jset_entry *);
332 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
333 					struct jset_entry *, unsigned long);
334 
335 void bch2_do_pending_node_rewrites(struct bch_fs *);
336 void bch2_free_pending_node_rewrites(struct bch_fs *);
337 
338 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
339 void bch2_fs_btree_interior_update_init_early(struct bch_fs *);
340 int bch2_fs_btree_interior_update_init(struct bch_fs *);
341 
342 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */
343