xref: /linux/fs/bcachefs/btree_update_interior.h (revision 335bbdf01d25517ae832ac1807fd8323c1f4f3b9)
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 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 *);
12 
13 #define BTREE_UPDATE_NODES_MAX		((BTREE_MAX_DEPTH - 2) * 2 + GC_MERGE_NODES)
14 
15 #define BTREE_UPDATE_JOURNAL_RES	(BTREE_UPDATE_NODES_MAX * (BKEY_BTREE_PTR_U64s_MAX + 1))
16 
17 /*
18  * Tracks an in progress split/rewrite of a btree node and the update to the
19  * parent node:
20  *
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.
24  *
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.
28  *
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.
33  *
34  */
35 struct btree_update {
36 	struct closure			cl;
37 	struct bch_fs			*c;
38 	u64				start_time;
39 
40 	struct list_head		list;
41 	struct list_head		unwritten_list;
42 
43 	/* What kind of update are we doing? */
44 	enum {
45 		BTREE_INTERIOR_NO_UPDATE,
46 		BTREE_INTERIOR_UPDATING_NODE,
47 		BTREE_INTERIOR_UPDATING_ROOT,
48 		BTREE_INTERIOR_UPDATING_AS,
49 	} mode;
50 
51 	unsigned			nodes_written:1;
52 	unsigned			took_gc_lock:1;
53 
54 	enum btree_id			btree_id;
55 	unsigned			update_level;
56 
57 	struct disk_reservation		disk_res;
58 	struct journal_preres		journal_preres;
59 
60 	/*
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:
67 	 */
68 	struct btree			*b;
69 	struct list_head		write_blocked_list;
70 
71 	/*
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:
76 	 */
77 	struct journal_entry_pin	journal;
78 
79 	/* Preallocated nodes we reserve when we start the update: */
80 	struct prealloc_nodes {
81 		struct btree		*b[BTREE_UPDATE_NODES_MAX];
82 		unsigned		nr;
83 	}				prealloc_nodes[2];
84 
85 	/* Nodes being freed: */
86 	struct keylist			old_keys;
87 	u64				_old_keys[BTREE_UPDATE_NODES_MAX *
88 						  BKEY_BTREE_PTR_U64s_MAX];
89 
90 	/* Nodes being added: */
91 	struct keylist			new_keys;
92 	u64				_new_keys[BTREE_UPDATE_NODES_MAX *
93 						  BKEY_BTREE_PTR_U64s_MAX];
94 
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;
98 
99 	struct btree			*old_nodes[BTREE_UPDATE_NODES_MAX];
100 	__le64				old_nodes_seq[BTREE_UPDATE_NODES_MAX];
101 	unsigned			nr_old_nodes;
102 
103 	open_bucket_idx_t		open_buckets[BTREE_UPDATE_NODES_MAX *
104 						     BCH_REPLICAS_MAX];
105 	open_bucket_idx_t		nr_open_buckets;
106 
107 	unsigned			journal_u64s;
108 	u64				journal_entries[BTREE_UPDATE_JOURNAL_RES];
109 
110 	/* Only here to reduce stack usage on recursive splits: */
111 	struct keylist			parent_keys;
112 	/*
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
116 	 */
117 	u64				inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
118 };
119 
120 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
121 						  struct btree_trans *,
122 						  struct btree *,
123 						  struct bkey_format);
124 
125 int bch2_btree_split_leaf(struct btree_trans *, struct btree_path *, unsigned);
126 
127 int __bch2_foreground_maybe_merge(struct btree_trans *, struct btree_path *,
128 				  unsigned, unsigned, enum btree_node_sibling);
129 
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)
134 {
135 	struct btree *b;
136 
137 	EBUG_ON(!btree_node_locked(path, level));
138 
139 	b = path->l[level].b;
140 	if (b->sib_u64s[sib] > trans->c->btree_foreground_merge_threshold)
141 		return 0;
142 
143 	return __bch2_foreground_maybe_merge(trans, path, level, flags, sib);
144 }
145 
146 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
147 					      struct btree_path *path,
148 					      unsigned level,
149 					      unsigned flags)
150 {
151 	return  bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
152 						    btree_prev_sib) ?:
153 		bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
154 						    btree_next_sib);
155 }
156 
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 *,
162 			       unsigned, bool);
163 int bch2_btree_node_update_key_get_iter(struct btree_trans *, struct btree *,
164 					struct bkey_i *, unsigned, bool);
165 
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);
168 
169 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
170 						     struct btree *b)
171 {
172 	unsigned depth = btree_node_root(c, b)->c.level + 1;
173 
174 	/*
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:
178 	 */
179 	if (depth < BTREE_MAX_DEPTH)
180 		return (depth - b->c.level) * 2 + 1;
181 	else
182 		return (depth - b->c.level) * 2 - 1;
183 }
184 
185 static inline void btree_node_reset_sib_u64s(struct btree *b)
186 {
187 	b->sib_u64s[0] = b->nr.live_u64s;
188 	b->sib_u64s[1] = b->nr.live_u64s;
189 }
190 
191 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
192 {
193 	return (void *) b->data + btree_bytes(c);
194 }
195 
196 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
197 							    struct btree *b)
198 {
199 	return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
200 }
201 
202 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
203 							  struct btree *b)
204 {
205 	return btree_data_end(c, b);
206 }
207 
208 static inline void *write_block(struct btree *b)
209 {
210 	return (void *) b->data + (b->written << 9);
211 }
212 
213 static inline bool __btree_addr_written(struct btree *b, void *p)
214 {
215 	return p < write_block(b);
216 }
217 
218 static inline bool bset_written(struct btree *b, struct bset *i)
219 {
220 	return __btree_addr_written(b, i);
221 }
222 
223 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
224 {
225 	return __btree_addr_written(b, k);
226 }
227 
228 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
229 						 struct btree *b,
230 						 void *end)
231 {
232 	ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
233 		b->whiteout_u64s;
234 	ssize_t total = c->opts.btree_node_size >> 3;
235 
236 	/* Always leave one extra u64 for bch2_varint_decode: */
237 	used++;
238 
239 	return total - used;
240 }
241 
242 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
243 						   struct btree *b)
244 {
245 	ssize_t remaining = __bch_btree_u64s_remaining(c, b,
246 				btree_bkey_last(b, bset_tree_last(b)));
247 
248 	BUG_ON(remaining < 0);
249 
250 	if (bset_written(b, btree_bset_last(b)))
251 		return 0;
252 
253 	return remaining;
254 }
255 
256 #define BTREE_WRITE_SET_U64s_BITS	9
257 
258 static inline unsigned btree_write_set_buffer(struct btree *b)
259 {
260 	/*
261 	 * Could buffer up larger amounts of keys for btrees with larger keys,
262 	 * pending benchmarking:
263 	 */
264 	return 8 << BTREE_WRITE_SET_U64s_BITS;
265 }
266 
267 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
268 						     struct btree *b)
269 {
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);
275 
276 	if (unlikely(bset_written(b, bset(b, t)))) {
277 		if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
278 			return bne;
279 	} else {
280 		if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
281 		    remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
282 			return bne;
283 	}
284 
285 	return NULL;
286 }
287 
288 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
289 				 struct bpos pos)
290 {
291 	struct bkey_packed k;
292 
293 	BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
294 	EBUG_ON(btree_node_just_written(b));
295 
296 	if (!bkey_pack_pos(&k, pos, b)) {
297 		struct bkey *u = (void *) &k;
298 
299 		bkey_init(u);
300 		u->p = pos;
301 	}
302 
303 	k.needs_whiteout = true;
304 
305 	b->whiteout_u64s += k.u64s;
306 	bkey_p_copy(unwritten_whiteouts_start(c, b), &k);
307 }
308 
309 /*
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)
312  */
313 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
314 					       struct btree *b, unsigned u64s)
315 {
316 	if (unlikely(btree_node_need_rewrite(b)))
317 		return false;
318 
319 	return u64s <= bch_btree_keys_u64s_remaining(c, b);
320 }
321 
322 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
323 
324 bool bch2_btree_interior_updates_flush(struct bch_fs *);
325 
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 *, unsigned long);
329 
330 void bch2_do_pending_node_rewrites(struct bch_fs *);
331 void bch2_free_pending_node_rewrites(struct bch_fs *);
332 
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 *);
336 
337 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */
338