git.sesse.net Git - bcachefs-tools-debian/blob - 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
  39         struct list_head                list;
  40         struct list_head                unwritten_list;
  41
  42         /* What kind of update are we doing? */
  43         enum {
  44                 BTREE_INTERIOR_NO_UPDATE,
  45                 BTREE_INTERIOR_UPDATING_NODE,
  46                 BTREE_INTERIOR_UPDATING_ROOT,
  47                 BTREE_INTERIOR_UPDATING_AS,
  48         } mode;
  49
  50         unsigned                        nodes_written:1;
  51         unsigned                        took_gc_lock:1;
  52
  53         enum btree_id                   btree_id;
  54
  55         struct disk_reservation         disk_res;
  56         struct journal_preres           journal_preres;
  57
  58         /*
  59          * BTREE_INTERIOR_UPDATING_NODE:
  60          * The update that made the new nodes visible was a regular update to an
  61          * existing interior node - @b. We can't write out the update to @b
  62          * until the new nodes we created are finished writing, so we block @b
  63          * from writing by putting this btree_interior update on the
  64          * @b->write_blocked list with @write_blocked_list:
  65          */
  66         struct btree                    *b;
  67         struct list_head                write_blocked_list;
  68
  69         /*
  70          * We may be freeing nodes that were dirty, and thus had journal entries
  71          * pinned: we need to transfer the oldest of those pins to the
  72          * btree_update operation, and release it when the new node(s)
  73          * are all persistent and reachable:
  74          */
  75         struct journal_entry_pin        journal;
  76
  77         /* Preallocated nodes we reserve when we start the update: */
  78         struct btree                    *prealloc_nodes[BTREE_UPDATE_NODES_MAX];
  79         unsigned                        nr_prealloc_nodes;
  80
  81         /* Nodes being freed: */
  82         struct keylist                  old_keys;
  83         u64                             _old_keys[BTREE_UPDATE_NODES_MAX *
  84                                                   BKEY_BTREE_PTR_VAL_U64s_MAX];
  85
  86         /* Nodes being added: */
  87         struct keylist                  new_keys;
  88         u64                             _new_keys[BTREE_UPDATE_NODES_MAX *
  89                                                   BKEY_BTREE_PTR_VAL_U64s_MAX];
  90
  91         /* New nodes, that will be made reachable by this update: */
  92         struct btree                    *new_nodes[BTREE_UPDATE_NODES_MAX];
  93         unsigned                        nr_new_nodes;
  94
  95         struct btree                    *old_nodes[BTREE_UPDATE_NODES_MAX];
  96         __le64                          old_nodes_seq[BTREE_UPDATE_NODES_MAX];
  97         unsigned                        nr_old_nodes;
  98
  99         open_bucket_idx_t               open_buckets[BTREE_UPDATE_NODES_MAX *
 100                                                      BCH_REPLICAS_MAX];
 101         open_bucket_idx_t               nr_open_buckets;
 102
 103         unsigned                        journal_u64s;
 104         u64                             journal_entries[BTREE_UPDATE_JOURNAL_RES];
 105
 106         /* Only here to reduce stack usage on recursive splits: */
 107         struct keylist                  parent_keys;
 108         /*
 109          * Enough room for btree_split's keys without realloc - btree node
 110          * pointers never have crc/compression info, so we only need to acount
 111          * for the pointers for three keys
 112          */
 113         u64                             inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
 114 };
 115
 116 void bch2_btree_node_free_inmem(struct bch_fs *, struct btree *,
 117                                 struct btree_iter *);
 118 void bch2_btree_node_free_never_inserted(struct bch_fs *, struct btree *);
 119
 120 void bch2_btree_update_get_open_buckets(struct btree_update *, struct btree *);
 121
 122 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
 123                                                   struct btree *,
 124                                                   struct bkey_format);
 125
 126 void bch2_btree_update_done(struct btree_update *);
 127 struct btree_update *
 128 bch2_btree_update_start(struct btree_iter *, unsigned, unsigned, unsigned);
 129
 130 void bch2_btree_interior_update_will_free_node(struct btree_update *,
 131                                                struct btree *);
 132 void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
 133
 134 int bch2_btree_split_leaf(struct btree_trans *, struct btree_iter *, unsigned);
 135
 136 int __bch2_foreground_maybe_merge(struct btree_trans *, struct btree_iter *,
 137                                   unsigned, unsigned, enum btree_node_sibling);
 138
 139 static inline int bch2_foreground_maybe_merge_sibling(struct btree_trans *trans,
 140                                         struct btree_iter *iter,
 141                                         unsigned level, unsigned flags,
 142                                         enum btree_node_sibling sib)
 143 {
 144         struct btree *b;
 145
 146         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 147                 return 0;
 148
 149         if (!bch2_btree_node_relock(iter, level))
 150                 return 0;
 151
 152         b = iter->l[level].b;
 153         if (b->sib_u64s[sib] > trans->c->btree_foreground_merge_threshold)
 154                 return 0;
 155
 156         return __bch2_foreground_maybe_merge(trans, iter, level, flags, sib);
 157 }
 158
 159 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
 160                                               struct btree_iter *iter,
 161                                               unsigned level,
 162                                               unsigned flags)
 163 {
 164         return  bch2_foreground_maybe_merge_sibling(trans, iter, level, flags,
 165                                                     btree_prev_sib) ?:
 166                 bch2_foreground_maybe_merge_sibling(trans, iter, level, flags,
 167                                                     btree_next_sib);
 168 }
 169
 170 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
 171 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
 172
 173 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
 174                                                      struct btree *b)
 175 {
 176         unsigned depth = btree_node_root(c, b)->c.level + 1;
 177
 178         /*
 179          * Number of nodes we might have to allocate in a worst case btree
 180          * split operation - we split all the way up to the root, then allocate
 181          * a new root, unless we're already at max depth:
 182          */
 183         if (depth < BTREE_MAX_DEPTH)
 184                 return (depth - b->c.level) * 2 + 1;
 185         else
 186                 return (depth - b->c.level) * 2 - 1;
 187 }
 188
 189 static inline void btree_node_reset_sib_u64s(struct btree *b)
 190 {
 191         b->sib_u64s[0] = b->nr.live_u64s;
 192         b->sib_u64s[1] = b->nr.live_u64s;
 193 }
 194
 195 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
 196 {
 197         return (void *) b->data + btree_bytes(c);
 198 }
 199
 200 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
 201                                                             struct btree *b)
 202 {
 203         return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
 204 }
 205
 206 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
 207                                                           struct btree *b)
 208 {
 209         return btree_data_end(c, b);
 210 }
 211
 212 static inline void *write_block(struct btree *b)
 213 {
 214         return (void *) b->data + (b->written << 9);
 215 }
 216
 217 static inline bool __btree_addr_written(struct btree *b, void *p)
 218 {
 219         return p < write_block(b);
 220 }
 221
 222 static inline bool bset_written(struct btree *b, struct bset *i)
 223 {
 224         return __btree_addr_written(b, i);
 225 }
 226
 227 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
 228 {
 229         return __btree_addr_written(b, k);
 230 }
 231
 232 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
 233                                                  struct btree *b,
 234                                                  void *end)
 235 {
 236         ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
 237                 b->whiteout_u64s;
 238         ssize_t total = c->opts.btree_node_size << 6;
 239
 240         /* Always leave one extra u64 for bch2_varint_decode: */
 241         used++;
 242
 243         return total - used;
 244 }
 245
 246 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
 247                                                    struct btree *b)
 248 {
 249         ssize_t remaining = __bch_btree_u64s_remaining(c, b,
 250                                 btree_bkey_last(b, bset_tree_last(b)));
 251
 252         BUG_ON(remaining < 0);
 253
 254         if (bset_written(b, btree_bset_last(b)))
 255                 return 0;
 256
 257         return remaining;
 258 }
 259
 260 #define BTREE_WRITE_SET_U64s_BITS       9
 261
 262 static inline unsigned btree_write_set_buffer(struct btree *b)
 263 {
 264         /*
 265          * Could buffer up larger amounts of keys for btrees with larger keys,
 266          * pending benchmarking:
 267          */
 268         return 8 << BTREE_WRITE_SET_U64s_BITS;
 269 }
 270
 271 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 272                                                      struct btree *b)
 273 {
 274         struct bset_tree *t = bset_tree_last(b);
 275         struct btree_node_entry *bne = max(write_block(b),
 276                         (void *) btree_bkey_last(b, bset_tree_last(b)));
 277         ssize_t remaining_space =
 278                 __bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
 279
 280         if (unlikely(bset_written(b, bset(b, t)))) {
 281                 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
 282                         return bne;
 283         } else {
 284                 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
 285                     remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
 286                         return bne;
 287         }
 288
 289         return NULL;
 290 }
 291
 292 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
 293                                  struct bpos pos)
 294 {
 295         struct bkey_packed k;
 296
 297         BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
 298
 299         if (!bkey_pack_pos(&k, pos, b)) {
 300                 struct bkey *u = (void *) &k;
 301
 302                 bkey_init(u);
 303                 u->p = pos;
 304         }
 305
 306         k.needs_whiteout = true;
 307
 308         b->whiteout_u64s += k.u64s;
 309         bkey_copy(unwritten_whiteouts_start(c, b), &k);
 310 }
 311
 312 /*
 313  * write lock must be held on @b (else the dirty bset that we were going to
 314  * insert into could be written out from under us)
 315  */
 316 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
 317                                                struct btree *b, unsigned u64s)
 318 {
 319         if (unlikely(btree_node_need_rewrite(b)))
 320                 return false;
 321
 322         return u64s <= bch_btree_keys_u64s_remaining(c, b);
 323 }
 324
 325 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
 326
 327 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *);
 328
 329 void bch2_journal_entries_to_btree_roots(struct bch_fs *, struct jset *);
 330 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
 331                                         struct jset_entry *, struct jset_entry *);
 332
 333 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
 334 int bch2_fs_btree_interior_update_init(struct bch_fs *);
 335
 336 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */