fs/bcachefs/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 #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_watermark              watermark;
  56         unsigned                        nodes_written:1;
  57         unsigned                        took_gc_lock:1;
  58
  59         enum btree_id                   btree_id;
  60         unsigned                        update_level;
  61
  62         struct disk_reservation         disk_res;
  63
  64         /*
  65          * BTREE_UPDATE_node:
  66          * The update that made the new nodes visible was a regular update to an
  67          * existing interior node - @b. We can't write out the update to @b
  68          * until the new nodes we created are finished writing, so we block @b
  69          * from writing by putting this btree_interior update on the
  70          * @b->write_blocked list with @write_blocked_list:
  71          */
  72         struct btree                    *b;
  73         struct list_head                write_blocked_list;
  74
  75         /*
  76          * We may be freeing nodes that were dirty, and thus had journal entries
  77          * pinned: we need to transfer the oldest of those pins to the
  78          * btree_update operation, and release it when the new node(s)
  79          * are all persistent and reachable:
  80          */
  81         struct journal_entry_pin        journal;
  82
  83         /* Preallocated nodes we reserve when we start the update: */
  84         struct prealloc_nodes {
  85                 struct btree            *b[BTREE_UPDATE_NODES_MAX];
  86                 unsigned                nr;
  87         }                               prealloc_nodes[2];
  88
  89         /* Nodes being freed: */
  90         struct keylist                  old_keys;
  91         u64                             _old_keys[BTREE_UPDATE_NODES_MAX *
  92                                                   BKEY_BTREE_PTR_U64s_MAX];
  93
  94         /* Nodes being added: */
  95         struct keylist                  new_keys;
  96         u64                             _new_keys[BTREE_UPDATE_NODES_MAX *
  97                                                   BKEY_BTREE_PTR_U64s_MAX];
  98
  99         /* New nodes, that will be made reachable by this update: */
 100         struct btree                    *new_nodes[BTREE_UPDATE_NODES_MAX];
 101         unsigned                        nr_new_nodes;
 102
 103         struct btree                    *old_nodes[BTREE_UPDATE_NODES_MAX];
 104         __le64                          old_nodes_seq[BTREE_UPDATE_NODES_MAX];
 105         unsigned                        nr_old_nodes;
 106
 107         open_bucket_idx_t               open_buckets[BTREE_UPDATE_NODES_MAX *
 108                                                      BCH_REPLICAS_MAX];
 109         open_bucket_idx_t               nr_open_buckets;
 110
 111         unsigned                        journal_u64s;
 112         u64                             journal_entries[BTREE_UPDATE_JOURNAL_RES];
 113
 114         /* Only here to reduce stack usage on recursive splits: */
 115         struct keylist                  parent_keys;
 116         /*
 117          * Enough room for btree_split's keys without realloc - btree node
 118          * pointers never have crc/compression info, so we only need to acount
 119          * for the pointers for three keys
 120          */
 121         u64                             inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
 122 };
 123
 124 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
 125                                                   struct btree_trans *,
 126                                                   struct btree *,
 127                                                   struct bkey_format);
 128
 129 int bch2_btree_split_leaf(struct btree_trans *, btree_path_idx_t, unsigned);
 130
 131 int bch2_btree_increase_depth(struct btree_trans *, btree_path_idx_t, unsigned);
 132
 133 int __bch2_foreground_maybe_merge(struct btree_trans *, btree_path_idx_t,
 134                                   unsigned, unsigned, enum btree_node_sibling);
 135
 136 static inline int bch2_foreground_maybe_merge_sibling(struct btree_trans *trans,
 137                                         btree_path_idx_t path_idx,
 138                                         unsigned level, unsigned flags,
 139                                         enum btree_node_sibling sib)
 140 {
 141         struct btree_path *path = trans->paths + path_idx;
 142         struct btree *b;
 143
 144         EBUG_ON(!btree_node_locked(path, level));
 145
 146         b = path->l[level].b;
 147         if (b->sib_u64s[sib] > trans->c->btree_foreground_merge_threshold)
 148                 return 0;
 149
 150         return __bch2_foreground_maybe_merge(trans, path_idx, level, flags, sib);
 151 }
 152
 153 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
 154                                               btree_path_idx_t path,
 155                                               unsigned level,
 156                                               unsigned flags)
 157 {
 158         return  bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
 159                                                     btree_prev_sib) ?:
 160                 bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
 161                                                     btree_next_sib);
 162 }
 163
 164 int bch2_btree_node_rewrite(struct btree_trans *, struct btree_iter *,
 165                             struct btree *, unsigned);
 166 void bch2_btree_node_rewrite_async(struct bch_fs *, struct btree *);
 167 int bch2_btree_node_update_key(struct btree_trans *, struct btree_iter *,
 168                                struct btree *, struct bkey_i *,
 169                                unsigned, bool);
 170 int bch2_btree_node_update_key_get_iter(struct btree_trans *, struct btree *,
 171                                         struct bkey_i *, unsigned, bool);
 172
 173 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
 174 void bch2_btree_root_alloc_fake(struct bch_fs *, enum btree_id, unsigned);
 175
 176 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
 177                                                      struct btree *b)
 178 {
 179         unsigned depth = btree_node_root(c, b)->c.level + 1;
 180
 181         /*
 182          * Number of nodes we might have to allocate in a worst case btree
 183          * split operation - we split all the way up to the root, then allocate
 184          * a new root, unless we're already at max depth:
 185          */
 186         if (depth < BTREE_MAX_DEPTH)
 187                 return (depth - b->c.level) * 2 + 1;
 188         else
 189                 return (depth - b->c.level) * 2 - 1;
 190 }
 191
 192 static inline void btree_node_reset_sib_u64s(struct btree *b)
 193 {
 194         b->sib_u64s[0] = b->nr.live_u64s;
 195         b->sib_u64s[1] = b->nr.live_u64s;
 196 }
 197
 198 static inline void *btree_data_end(struct btree *b)
 199 {
 200         return (void *) b->data + btree_buf_bytes(b);
 201 }
 202
 203 static inline struct bkey_packed *unwritten_whiteouts_start(struct btree *b)
 204 {
 205         return (void *) ((u64 *) btree_data_end(b) - b->whiteout_u64s);
 206 }
 207
 208 static inline struct bkey_packed *unwritten_whiteouts_end(struct btree *b)
 209 {
 210         return btree_data_end(b);
 211 }
 212
 213 static inline void *write_block(struct btree *b)
 214 {
 215         return (void *) b->data + (b->written << 9);
 216 }
 217
 218 static inline bool __btree_addr_written(struct btree *b, void *p)
 219 {
 220         return p < write_block(b);
 221 }
 222
 223 static inline bool bset_written(struct btree *b, struct bset *i)
 224 {
 225         return __btree_addr_written(b, i);
 226 }
 227
 228 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
 229 {
 230         return __btree_addr_written(b, k);
 231 }
 232
 233 static inline ssize_t __bch2_btree_u64s_remaining(struct btree *b, void *end)
 234 {
 235         ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
 236                 b->whiteout_u64s;
 237         ssize_t total = btree_buf_bytes(b) >> 3;
 238
 239         /* Always leave one extra u64 for bch2_varint_decode: */
 240         used++;
 241
 242         return total - used;
 243 }
 244
 245 static inline size_t bch2_btree_keys_u64s_remaining(struct btree *b)
 246 {
 247         ssize_t remaining = __bch2_btree_u64s_remaining(b,
 248                                 btree_bkey_last(b, bset_tree_last(b)));
 249
 250         BUG_ON(remaining < 0);
 251
 252         if (bset_written(b, btree_bset_last(b)))
 253                 return 0;
 254
 255         return remaining;
 256 }
 257
 258 #define BTREE_WRITE_SET_U64s_BITS       9
 259
 260 static inline unsigned btree_write_set_buffer(struct btree *b)
 261 {
 262         /*
 263          * Could buffer up larger amounts of keys for btrees with larger keys,
 264          * pending benchmarking:
 265          */
 266         return 8 << BTREE_WRITE_SET_U64s_BITS;
 267 }
 268
 269 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c, struct btree *b)
 270 {
 271         struct bset_tree *t = bset_tree_last(b);
 272         struct btree_node_entry *bne = max(write_block(b),
 273                         (void *) btree_bkey_last(b, bset_tree_last(b)));
 274         ssize_t remaining_space =
 275                 __bch2_btree_u64s_remaining(b, bne->keys.start);
 276
 277         if (unlikely(bset_written(b, bset(b, t)))) {
 278                 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
 279                         return bne;
 280         } else {
 281                 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
 282                     remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
 283                         return bne;
 284         }
 285
 286         return NULL;
 287 }
 288
 289 static inline void push_whiteout(struct btree *b, struct bpos pos)
 290 {
 291         struct bkey_packed k;
 292
 293         BUG_ON(bch2_btree_keys_u64s_remaining(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(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 btree *b, unsigned u64s)
 314 {
 315         if (unlikely(btree_node_need_rewrite(b)))
 316                 return false;
 317
 318         return u64s <= bch2_btree_keys_u64s_remaining(b);
 319 }
 320
 321 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
 322
 323 bool bch2_btree_interior_updates_flush(struct bch_fs *);
 324
 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);
 328
 329 void bch2_do_pending_node_rewrites(struct bch_fs *);
 330 void bch2_free_pending_node_rewrites(struct bch_fs *);
 331
 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 *);
 335
 336 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */