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 #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 /*
  14  * Tracks an in progress split/rewrite of a btree node and the update to the
  15  * parent node:
  16  *
  17  * When we split/rewrite a node, we do all the updates in memory without
  18  * waiting for any writes to complete - we allocate the new node(s) and update
  19  * the parent node, possibly recursively up to the root.
  20  *
  21  * The end result is that we have one or more new nodes being written -
  22  * possibly several, if there were multiple splits - and then a write (updating
  23  * an interior node) which will make all these new nodes visible.
  24  *
  25  * Additionally, as we split/rewrite nodes we free the old nodes - but the old
  26  * nodes can't be freed (their space on disk can't be reclaimed) until the
  27  * update to the interior node that makes the new node visible completes -
  28  * until then, the old nodes are still reachable on disk.
  29  *
  30  */
  31 struct btree_update {
  32         struct closure                  cl;
  33         struct bch_fs                   *c;
  34         u64                             start_time;
  35
  36         struct list_head                list;
  37         struct list_head                unwritten_list;
  38
  39         /* What kind of update are we doing? */
  40         enum {
  41                 BTREE_INTERIOR_NO_UPDATE,
  42                 BTREE_INTERIOR_UPDATING_NODE,
  43                 BTREE_INTERIOR_UPDATING_ROOT,
  44                 BTREE_INTERIOR_UPDATING_AS,
  45         } mode;
  46
  47         unsigned                        nodes_written:1;
  48         unsigned                        took_gc_lock:1;
  49
  50         enum btree_id                   btree_id;
  51         unsigned                        update_level;
  52
  53         struct disk_reservation         disk_res;
  54
  55         /*
  56          * BTREE_INTERIOR_UPDATING_NODE:
  57          * The update that made the new nodes visible was a regular update to an
  58          * existing interior node - @b. We can't write out the update to @b
  59          * until the new nodes we created are finished writing, so we block @b
  60          * from writing by putting this btree_interior update on the
  61          * @b->write_blocked list with @write_blocked_list:
  62          */
  63         struct btree                    *b;
  64         struct list_head                write_blocked_list;
  65
  66         /*
  67          * We may be freeing nodes that were dirty, and thus had journal entries
  68          * pinned: we need to transfer the oldest of those pins to the
  69          * btree_update operation, and release it when the new node(s)
  70          * are all persistent and reachable:
  71          */
  72         struct journal_entry_pin        journal;
  73
  74         /* Preallocated nodes we reserve when we start the update: */
  75         struct prealloc_nodes {
  76                 struct btree            *b[BTREE_UPDATE_NODES_MAX];
  77                 unsigned                nr;
  78         }                               prealloc_nodes[2];
  79
  80         /* Nodes being freed: */
  81         struct keylist                  old_keys;
  82         u64                             _old_keys[BTREE_UPDATE_NODES_MAX *
  83                                                   BKEY_BTREE_PTR_U64s_MAX];
  84
  85         /* Nodes being added: */
  86         struct keylist                  new_keys;
  87         u64                             _new_keys[BTREE_UPDATE_NODES_MAX *
  88                                                   BKEY_BTREE_PTR_U64s_MAX];
  89
  90         /* New nodes, that will be made reachable by this update: */
  91         struct btree                    *new_nodes[BTREE_UPDATE_NODES_MAX];
  92         unsigned                        nr_new_nodes;
  93
  94         struct btree                    *old_nodes[BTREE_UPDATE_NODES_MAX];
  95         __le64                          old_nodes_seq[BTREE_UPDATE_NODES_MAX];
  96         unsigned                        nr_old_nodes;
  97
  98         open_bucket_idx_t               open_buckets[BTREE_UPDATE_NODES_MAX *
  99                                                      BCH_REPLICAS_MAX];
 100         open_bucket_idx_t               nr_open_buckets;
 101
 102         unsigned                        journal_u64s;
 103         u64                             journal_entries[BTREE_UPDATE_JOURNAL_RES];
 104
 105         /* Only here to reduce stack usage on recursive splits: */
 106         struct keylist                  parent_keys;
 107         /*
 108          * Enough room for btree_split's keys without realloc - btree node
 109          * pointers never have crc/compression info, so we only need to acount
 110          * for the pointers for three keys
 111          */
 112         u64                             inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
 113 };
 114
 115 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
 116                                                   struct btree_trans *,
 117                                                   struct btree *,
 118                                                   struct bkey_format);
 119
 120 int bch2_btree_split_leaf(struct btree_trans *, btree_path_idx_t, unsigned);
 121
 122 int __bch2_foreground_maybe_merge(struct btree_trans *, btree_path_idx_t,
 123                                   unsigned, unsigned, enum btree_node_sibling);
 124
 125 static inline int bch2_foreground_maybe_merge_sibling(struct btree_trans *trans,
 126                                         btree_path_idx_t path_idx,
 127                                         unsigned level, unsigned flags,
 128                                         enum btree_node_sibling sib)
 129 {
 130         struct btree_path *path = trans->paths + path_idx;
 131         struct btree *b;
 132
 133         EBUG_ON(!btree_node_locked(path, level));
 134
 135         b = path->l[level].b;
 136         if (b->sib_u64s[sib] > trans->c->btree_foreground_merge_threshold)
 137                 return 0;
 138
 139         return __bch2_foreground_maybe_merge(trans, path_idx, level, flags, sib);
 140 }
 141
 142 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
 143                                               btree_path_idx_t path,
 144                                               unsigned level,
 145                                               unsigned flags)
 146 {
 147         return  bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
 148                                                     btree_prev_sib) ?:
 149                 bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
 150                                                     btree_next_sib);
 151 }
 152
 153 int bch2_btree_node_rewrite(struct btree_trans *, struct btree_iter *,
 154                             struct btree *, unsigned);
 155 void bch2_btree_node_rewrite_async(struct bch_fs *, struct btree *);
 156 int bch2_btree_node_update_key(struct btree_trans *, struct btree_iter *,
 157                                struct btree *, struct bkey_i *,
 158                                unsigned, bool);
 159 int bch2_btree_node_update_key_get_iter(struct btree_trans *, struct btree *,
 160                                         struct bkey_i *, unsigned, bool);
 161
 162 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
 163 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
 164
 165 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
 166                                                      struct btree *b)
 167 {
 168         unsigned depth = btree_node_root(c, b)->c.level + 1;
 169
 170         /*
 171          * Number of nodes we might have to allocate in a worst case btree
 172          * split operation - we split all the way up to the root, then allocate
 173          * a new root, unless we're already at max depth:
 174          */
 175         if (depth < BTREE_MAX_DEPTH)
 176                 return (depth - b->c.level) * 2 + 1;
 177         else
 178                 return (depth - b->c.level) * 2 - 1;
 179 }
 180
 181 static inline void btree_node_reset_sib_u64s(struct btree *b)
 182 {
 183         b->sib_u64s[0] = b->nr.live_u64s;
 184         b->sib_u64s[1] = b->nr.live_u64s;
 185 }
 186
 187 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
 188 {
 189         return (void *) b->data + btree_bytes(c);
 190 }
 191
 192 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
 193                                                             struct btree *b)
 194 {
 195         return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
 196 }
 197
 198 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
 199                                                           struct btree *b)
 200 {
 201         return btree_data_end(c, b);
 202 }
 203
 204 static inline void *write_block(struct btree *b)
 205 {
 206         return (void *) b->data + (b->written << 9);
 207 }
 208
 209 static inline bool __btree_addr_written(struct btree *b, void *p)
 210 {
 211         return p < write_block(b);
 212 }
 213
 214 static inline bool bset_written(struct btree *b, struct bset *i)
 215 {
 216         return __btree_addr_written(b, i);
 217 }
 218
 219 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
 220 {
 221         return __btree_addr_written(b, k);
 222 }
 223
 224 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
 225                                                  struct btree *b,
 226                                                  void *end)
 227 {
 228         ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
 229                 b->whiteout_u64s;
 230         ssize_t total = c->opts.btree_node_size >> 3;
 231
 232         /* Always leave one extra u64 for bch2_varint_decode: */
 233         used++;
 234
 235         return total - used;
 236 }
 237
 238 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
 239                                                    struct btree *b)
 240 {
 241         ssize_t remaining = __bch_btree_u64s_remaining(c, b,
 242                                 btree_bkey_last(b, bset_tree_last(b)));
 243
 244         BUG_ON(remaining < 0);
 245
 246         if (bset_written(b, btree_bset_last(b)))
 247                 return 0;
 248
 249         return remaining;
 250 }
 251
 252 #define BTREE_WRITE_SET_U64s_BITS       9
 253
 254 static inline unsigned btree_write_set_buffer(struct btree *b)
 255 {
 256         /*
 257          * Could buffer up larger amounts of keys for btrees with larger keys,
 258          * pending benchmarking:
 259          */
 260         return 8 << BTREE_WRITE_SET_U64s_BITS;
 261 }
 262
 263 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 264                                                      struct btree *b)
 265 {
 266         struct bset_tree *t = bset_tree_last(b);
 267         struct btree_node_entry *bne = max(write_block(b),
 268                         (void *) btree_bkey_last(b, bset_tree_last(b)));
 269         ssize_t remaining_space =
 270                 __bch_btree_u64s_remaining(c, b, bne->keys.start);
 271
 272         if (unlikely(bset_written(b, bset(b, t)))) {
 273                 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
 274                         return bne;
 275         } else {
 276                 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
 277                     remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
 278                         return bne;
 279         }
 280
 281         return NULL;
 282 }
 283
 284 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
 285                                  struct bpos pos)
 286 {
 287         struct bkey_packed k;
 288
 289         BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
 290         EBUG_ON(btree_node_just_written(b));
 291
 292         if (!bkey_pack_pos(&k, pos, b)) {
 293                 struct bkey *u = (void *) &k;
 294
 295                 bkey_init(u);
 296                 u->p = pos;
 297         }
 298
 299         k.needs_whiteout = true;
 300
 301         b->whiteout_u64s += k.u64s;
 302         bkey_p_copy(unwritten_whiteouts_start(c, b), &k);
 303 }
 304
 305 /*
 306  * write lock must be held on @b (else the dirty bset that we were going to
 307  * insert into could be written out from under us)
 308  */
 309 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
 310                                                struct btree *b, unsigned u64s)
 311 {
 312         if (unlikely(btree_node_need_rewrite(b)))
 313                 return false;
 314
 315         return u64s <= bch_btree_keys_u64s_remaining(c, b);
 316 }
 317
 318 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
 319
 320 bool bch2_btree_interior_updates_flush(struct bch_fs *);
 321
 322 void bch2_journal_entry_to_btree_root(struct bch_fs *, struct jset_entry *);
 323 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
 324                                         struct jset_entry *, unsigned long);
 325
 326 void bch2_do_pending_node_rewrites(struct bch_fs *);
 327 void bch2_free_pending_node_rewrites(struct bch_fs *);
 328
 329 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
 330 void bch2_fs_btree_interior_update_init_early(struct bch_fs *);
 331 int bch2_fs_btree_interior_update_init(struct bch_fs *);
 332
 333 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */