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         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 *,
 121                                                   struct bkey_format);
 122
 123 int bch2_btree_split_leaf(struct btree_trans *, struct btree_path *, unsigned);
 124
 125 int __bch2_foreground_maybe_merge(struct btree_trans *, struct btree_path *,
 126                                   unsigned, unsigned, enum btree_node_sibling);
 127
 128 static inline int bch2_foreground_maybe_merge_sibling(struct btree_trans *trans,
 129                                         struct btree_path *path,
 130                                         unsigned level, unsigned flags,
 131                                         enum btree_node_sibling sib)
 132 {
 133         struct btree *b;
 134
 135         EBUG_ON(!btree_node_locked(path, level));
 136
 137         b = path->l[level].b;
 138         if (b->sib_u64s[sib] > trans->c->btree_foreground_merge_threshold)
 139                 return 0;
 140
 141         return __bch2_foreground_maybe_merge(trans, path, level, flags, sib);
 142 }
 143
 144 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
 145                                               struct btree_path *path,
 146                                               unsigned level,
 147                                               unsigned flags)
 148 {
 149         return  bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
 150                                                     btree_prev_sib) ?:
 151                 bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
 152                                                     btree_next_sib);
 153 }
 154
 155 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
 156 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
 157
 158 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
 159                                                      struct btree *b)
 160 {
 161         unsigned depth = btree_node_root(c, b)->c.level + 1;
 162
 163         /*
 164          * Number of nodes we might have to allocate in a worst case btree
 165          * split operation - we split all the way up to the root, then allocate
 166          * a new root, unless we're already at max depth:
 167          */
 168         if (depth < BTREE_MAX_DEPTH)
 169                 return (depth - b->c.level) * 2 + 1;
 170         else
 171                 return (depth - b->c.level) * 2 - 1;
 172 }
 173
 174 static inline void btree_node_reset_sib_u64s(struct btree *b)
 175 {
 176         b->sib_u64s[0] = b->nr.live_u64s;
 177         b->sib_u64s[1] = b->nr.live_u64s;
 178 }
 179
 180 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
 181 {
 182         return (void *) b->data + btree_bytes(c);
 183 }
 184
 185 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
 186                                                             struct btree *b)
 187 {
 188         return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
 189 }
 190
 191 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
 192                                                           struct btree *b)
 193 {
 194         return btree_data_end(c, b);
 195 }
 196
 197 static inline void *write_block(struct btree *b)
 198 {
 199         return (void *) b->data + (b->written << 9);
 200 }
 201
 202 static inline bool __btree_addr_written(struct btree *b, void *p)
 203 {
 204         return p < write_block(b);
 205 }
 206
 207 static inline bool bset_written(struct btree *b, struct bset *i)
 208 {
 209         return __btree_addr_written(b, i);
 210 }
 211
 212 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
 213 {
 214         return __btree_addr_written(b, k);
 215 }
 216
 217 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
 218                                                  struct btree *b,
 219                                                  void *end)
 220 {
 221         ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
 222                 b->whiteout_u64s;
 223         ssize_t total = c->opts.btree_node_size >> 3;
 224
 225         /* Always leave one extra u64 for bch2_varint_decode: */
 226         used++;
 227
 228         return total - used;
 229 }
 230
 231 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
 232                                                    struct btree *b)
 233 {
 234         ssize_t remaining = __bch_btree_u64s_remaining(c, b,
 235                                 btree_bkey_last(b, bset_tree_last(b)));
 236
 237         BUG_ON(remaining < 0);
 238
 239         if (bset_written(b, btree_bset_last(b)))
 240                 return 0;
 241
 242         return remaining;
 243 }
 244
 245 #define BTREE_WRITE_SET_U64s_BITS       9
 246
 247 static inline unsigned btree_write_set_buffer(struct btree *b)
 248 {
 249         /*
 250          * Could buffer up larger amounts of keys for btrees with larger keys,
 251          * pending benchmarking:
 252          */
 253         return 8 << BTREE_WRITE_SET_U64s_BITS;
 254 }
 255
 256 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 257                                                      struct btree *b)
 258 {
 259         struct bset_tree *t = bset_tree_last(b);
 260         struct btree_node_entry *bne = max(write_block(b),
 261                         (void *) btree_bkey_last(b, bset_tree_last(b)));
 262         ssize_t remaining_space =
 263                 __bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
 264
 265         if (unlikely(bset_written(b, bset(b, t)))) {
 266                 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
 267                         return bne;
 268         } else {
 269                 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
 270                     remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
 271                         return bne;
 272         }
 273
 274         return NULL;
 275 }
 276
 277 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
 278                                  struct bpos pos)
 279 {
 280         struct bkey_packed k;
 281
 282         BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
 283
 284         if (!bkey_pack_pos(&k, pos, b)) {
 285                 struct bkey *u = (void *) &k;
 286
 287                 bkey_init(u);
 288                 u->p = pos;
 289         }
 290
 291         k.needs_whiteout = true;
 292
 293         b->whiteout_u64s += k.u64s;
 294         bkey_copy(unwritten_whiteouts_start(c, b), &k);
 295 }
 296
 297 /*
 298  * write lock must be held on @b (else the dirty bset that we were going to
 299  * insert into could be written out from under us)
 300  */
 301 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
 302                                                struct btree *b, unsigned u64s)
 303 {
 304         if (unlikely(btree_node_need_rewrite(b)))
 305                 return false;
 306
 307         return u64s <= bch_btree_keys_u64s_remaining(c, b);
 308 }
 309
 310 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
 311
 312 bool bch2_btree_interior_updates_flush(struct bch_fs *);
 313
 314 void bch2_journal_entries_to_btree_roots(struct bch_fs *, struct jset *);
 315 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
 316                                         struct jset_entry *, struct jset_entry *);
 317
 318 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
 319 int bch2_fs_btree_interior_update_init(struct bch_fs *);
 320
 321 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */