]> git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_interior.h
Update bcachefs sources to 176718966e bcachefs: Re-enable hash_redo_key()
[bcachefs-tools-debian] / libbcachefs / btree_update_interior.h
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
56         struct disk_reservation         disk_res;
57         struct journal_preres           journal_preres;
58
59         /*
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:
66          */
67         struct btree                    *b;
68         struct list_head                write_blocked_list;
69
70         /*
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:
75          */
76         struct journal_entry_pin        journal;
77
78         /* Preallocated nodes we reserve when we start the update: */
79         struct prealloc_nodes {
80                 struct btree            *b[BTREE_UPDATE_NODES_MAX];
81                 unsigned                nr;
82         }                               prealloc_nodes[2];
83
84         /* Nodes being freed: */
85         struct keylist                  old_keys;
86         u64                             _old_keys[BTREE_UPDATE_NODES_MAX *
87                                                   BKEY_BTREE_PTR_U64s_MAX];
88
89         /* Nodes being added: */
90         struct keylist                  new_keys;
91         u64                             _new_keys[BTREE_UPDATE_NODES_MAX *
92                                                   BKEY_BTREE_PTR_U64s_MAX];
93
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;
97
98         struct btree                    *old_nodes[BTREE_UPDATE_NODES_MAX];
99         __le64                          old_nodes_seq[BTREE_UPDATE_NODES_MAX];
100         unsigned                        nr_old_nodes;
101
102         open_bucket_idx_t               open_buckets[BTREE_UPDATE_NODES_MAX *
103                                                      BCH_REPLICAS_MAX];
104         open_bucket_idx_t               nr_open_buckets;
105
106         unsigned                        journal_u64s;
107         u64                             journal_entries[BTREE_UPDATE_JOURNAL_RES];
108
109         /* Only here to reduce stack usage on recursive splits: */
110         struct keylist                  parent_keys;
111         /*
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
115          */
116         u64                             inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
117 };
118
119 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
120                                                   struct btree_trans *,
121                                                   struct btree *,
122                                                   struct bkey_format);
123
124 int bch2_btree_split_leaf(struct btree_trans *, struct btree_path *, unsigned);
125
126 int __bch2_foreground_maybe_merge(struct btree_trans *, struct btree_path *,
127                                   unsigned, unsigned, enum btree_node_sibling);
128
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)
133 {
134         struct btree *b;
135
136         EBUG_ON(!btree_node_locked(path, level));
137
138         b = path->l[level].b;
139         if (b->sib_u64s[sib] > trans->c->btree_foreground_merge_threshold)
140                 return 0;
141
142         return __bch2_foreground_maybe_merge(trans, path, level, flags, sib);
143 }
144
145 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
146                                               struct btree_path *path,
147                                               unsigned level,
148                                               unsigned flags)
149 {
150         return  bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
151                                                     btree_prev_sib) ?:
152                 bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
153                                                     btree_next_sib);
154 }
155
156 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
157 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
158
159 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
160                                                      struct btree *b)
161 {
162         unsigned depth = btree_node_root(c, b)->c.level + 1;
163
164         /*
165          * Number of nodes we might have to allocate in a worst case btree
166          * split operation - we split all the way up to the root, then allocate
167          * a new root, unless we're already at max depth:
168          */
169         if (depth < BTREE_MAX_DEPTH)
170                 return (depth - b->c.level) * 2 + 1;
171         else
172                 return (depth - b->c.level) * 2 - 1;
173 }
174
175 static inline void btree_node_reset_sib_u64s(struct btree *b)
176 {
177         b->sib_u64s[0] = b->nr.live_u64s;
178         b->sib_u64s[1] = b->nr.live_u64s;
179 }
180
181 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
182 {
183         return (void *) b->data + btree_bytes(c);
184 }
185
186 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
187                                                             struct btree *b)
188 {
189         return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
190 }
191
192 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
193                                                           struct btree *b)
194 {
195         return btree_data_end(c, b);
196 }
197
198 static inline void *write_block(struct btree *b)
199 {
200         return (void *) b->data + (b->written << 9);
201 }
202
203 static inline bool __btree_addr_written(struct btree *b, void *p)
204 {
205         return p < write_block(b);
206 }
207
208 static inline bool bset_written(struct btree *b, struct bset *i)
209 {
210         return __btree_addr_written(b, i);
211 }
212
213 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
214 {
215         return __btree_addr_written(b, k);
216 }
217
218 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
219                                                  struct btree *b,
220                                                  void *end)
221 {
222         ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
223                 b->whiteout_u64s;
224         ssize_t total = c->opts.btree_node_size >> 3;
225
226         /* Always leave one extra u64 for bch2_varint_decode: */
227         used++;
228
229         return total - used;
230 }
231
232 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
233                                                    struct btree *b)
234 {
235         ssize_t remaining = __bch_btree_u64s_remaining(c, b,
236                                 btree_bkey_last(b, bset_tree_last(b)));
237
238         BUG_ON(remaining < 0);
239
240         if (bset_written(b, btree_bset_last(b)))
241                 return 0;
242
243         return remaining;
244 }
245
246 #define BTREE_WRITE_SET_U64s_BITS       9
247
248 static inline unsigned btree_write_set_buffer(struct btree *b)
249 {
250         /*
251          * Could buffer up larger amounts of keys for btrees with larger keys,
252          * pending benchmarking:
253          */
254         return 8 << BTREE_WRITE_SET_U64s_BITS;
255 }
256
257 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
258                                                      struct btree *b)
259 {
260         struct bset_tree *t = bset_tree_last(b);
261         struct btree_node_entry *bne = max(write_block(b),
262                         (void *) btree_bkey_last(b, bset_tree_last(b)));
263         ssize_t remaining_space =
264                 __bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
265
266         if (unlikely(bset_written(b, bset(b, t)))) {
267                 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
268                         return bne;
269         } else {
270                 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
271                     remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
272                         return bne;
273         }
274
275         return NULL;
276 }
277
278 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
279                                  struct bpos pos)
280 {
281         struct bkey_packed k;
282
283         BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
284
285         if (!bkey_pack_pos(&k, pos, b)) {
286                 struct bkey *u = (void *) &k;
287
288                 bkey_init(u);
289                 u->p = pos;
290         }
291
292         k.needs_whiteout = true;
293
294         b->whiteout_u64s += k.u64s;
295         bkey_copy(unwritten_whiteouts_start(c, b), &k);
296 }
297
298 /*
299  * write lock must be held on @b (else the dirty bset that we were going to
300  * insert into could be written out from under us)
301  */
302 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
303                                                struct btree *b, unsigned u64s)
304 {
305         if (unlikely(btree_node_need_rewrite(b)))
306                 return false;
307
308         return u64s <= bch_btree_keys_u64s_remaining(c, b);
309 }
310
311 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
312
313 bool bch2_btree_interior_updates_flush(struct bch_fs *);
314
315 void bch2_journal_entries_to_btree_roots(struct bch_fs *, struct jset *);
316 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
317                                         struct jset_entry *, struct jset_entry *);
318
319 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
320 int bch2_fs_btree_interior_update_init(struct bch_fs *);
321
322 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */