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 btree                    *prealloc_nodes[BTREE_UPDATE_NODES_MAX];
  80         unsigned                        nr_prealloc_nodes;
  81
  82         /* Nodes being freed: */
  83         struct keylist                  old_keys;
  84         u64                             _old_keys[BTREE_UPDATE_NODES_MAX *
  85                                                   BKEY_BTREE_PTR_U64s_MAX];
  86
  87         /* Nodes being added: */
  88         struct keylist                  new_keys;
  89         u64                             _new_keys[BTREE_UPDATE_NODES_MAX *
  90                                                   BKEY_BTREE_PTR_U64s_MAX];
  91
  92         /* New nodes, that will be made reachable by this update: */
  93         struct btree                    *new_nodes[BTREE_UPDATE_NODES_MAX];
  94         unsigned                        nr_new_nodes;
  95
  96         struct btree                    *old_nodes[BTREE_UPDATE_NODES_MAX];
  97         __le64                          old_nodes_seq[BTREE_UPDATE_NODES_MAX];
  98         unsigned                        nr_old_nodes;
  99
 100         open_bucket_idx_t               open_buckets[BTREE_UPDATE_NODES_MAX *
 101                                                      BCH_REPLICAS_MAX];
 102         open_bucket_idx_t               nr_open_buckets;
 103
 104         unsigned                        journal_u64s;
 105         u64                             journal_entries[BTREE_UPDATE_JOURNAL_RES];
 106
 107         /* Only here to reduce stack usage on recursive splits: */
 108         struct keylist                  parent_keys;
 109         /*
 110          * Enough room for btree_split's keys without realloc - btree node
 111          * pointers never have crc/compression info, so we only need to acount
 112          * for the pointers for three keys
 113          */
 114         u64                             inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
 115 };
 116
 117 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
 118                                                   struct btree *,
 119                                                   struct bkey_format);
 120
 121 int bch2_btree_split_leaf(struct btree_trans *, struct btree_path *, unsigned);
 122
 123 int __bch2_foreground_maybe_merge(struct btree_trans *, struct btree_path *,
 124                                   unsigned, unsigned, enum btree_node_sibling);
 125
 126 static inline int bch2_foreground_maybe_merge_sibling(struct btree_trans *trans,
 127                                         struct btree_path *path,
 128                                         unsigned level, unsigned flags,
 129                                         enum btree_node_sibling sib)
 130 {
 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, level, flags, sib);
 140 }
 141
 142 static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
 143                                               struct btree_path *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 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
 154 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
 155
 156 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
 157                                                      struct btree *b)
 158 {
 159         unsigned depth = btree_node_root(c, b)->c.level + 1;
 160
 161         /*
 162          * Number of nodes we might have to allocate in a worst case btree
 163          * split operation - we split all the way up to the root, then allocate
 164          * a new root, unless we're already at max depth:
 165          */
 166         if (depth < BTREE_MAX_DEPTH)
 167                 return (depth - b->c.level) * 2 + 1;
 168         else
 169                 return (depth - b->c.level) * 2 - 1;
 170 }
 171
 172 static inline void btree_node_reset_sib_u64s(struct btree *b)
 173 {
 174         b->sib_u64s[0] = b->nr.live_u64s;
 175         b->sib_u64s[1] = b->nr.live_u64s;
 176 }
 177
 178 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
 179 {
 180         return (void *) b->data + btree_bytes(c);
 181 }
 182
 183 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
 184                                                             struct btree *b)
 185 {
 186         return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
 187 }
 188
 189 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
 190                                                           struct btree *b)
 191 {
 192         return btree_data_end(c, b);
 193 }
 194
 195 static inline void *write_block(struct btree *b)
 196 {
 197         return (void *) b->data + (b->written << 9);
 198 }
 199
 200 static inline bool __btree_addr_written(struct btree *b, void *p)
 201 {
 202         return p < write_block(b);
 203 }
 204
 205 static inline bool bset_written(struct btree *b, struct bset *i)
 206 {
 207         return __btree_addr_written(b, i);
 208 }
 209
 210 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
 211 {
 212         return __btree_addr_written(b, k);
 213 }
 214
 215 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
 216                                                  struct btree *b,
 217                                                  void *end)
 218 {
 219         ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
 220                 b->whiteout_u64s;
 221         ssize_t total = c->opts.btree_node_size >> 3;
 222
 223         /* Always leave one extra u64 for bch2_varint_decode: */
 224         used++;
 225
 226         return total - used;
 227 }
 228
 229 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
 230                                                    struct btree *b)
 231 {
 232         ssize_t remaining = __bch_btree_u64s_remaining(c, b,
 233                                 btree_bkey_last(b, bset_tree_last(b)));
 234
 235         BUG_ON(remaining < 0);
 236
 237         if (bset_written(b, btree_bset_last(b)))
 238                 return 0;
 239
 240         return remaining;
 241 }
 242
 243 #define BTREE_WRITE_SET_U64s_BITS       9
 244
 245 static inline unsigned btree_write_set_buffer(struct btree *b)
 246 {
 247         /*
 248          * Could buffer up larger amounts of keys for btrees with larger keys,
 249          * pending benchmarking:
 250          */
 251         return 8 << BTREE_WRITE_SET_U64s_BITS;
 252 }
 253
 254 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 255                                                      struct btree *b)
 256 {
 257         struct bset_tree *t = bset_tree_last(b);
 258         struct btree_node_entry *bne = max(write_block(b),
 259                         (void *) btree_bkey_last(b, bset_tree_last(b)));
 260         ssize_t remaining_space =
 261                 __bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
 262
 263         if (unlikely(bset_written(b, bset(b, t)))) {
 264                 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
 265                         return bne;
 266         } else {
 267                 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
 268                     remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
 269                         return bne;
 270         }
 271
 272         return NULL;
 273 }
 274
 275 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
 276                                  struct bpos pos)
 277 {
 278         struct bkey_packed k;
 279
 280         BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
 281
 282         if (!bkey_pack_pos(&k, pos, b)) {
 283                 struct bkey *u = (void *) &k;
 284
 285                 bkey_init(u);
 286                 u->p = pos;
 287         }
 288
 289         k.needs_whiteout = true;
 290
 291         b->whiteout_u64s += k.u64s;
 292         bkey_copy(unwritten_whiteouts_start(c, b), &k);
 293 }
 294
 295 /*
 296  * write lock must be held on @b (else the dirty bset that we were going to
 297  * insert into could be written out from under us)
 298  */
 299 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
 300                                                struct btree *b, unsigned u64s)
 301 {
 302         if (unlikely(btree_node_need_rewrite(b)))
 303                 return false;
 304
 305         return u64s <= bch_btree_keys_u64s_remaining(c, b);
 306 }
 307
 308 void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
 309
 310 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *);
 311
 312 void bch2_journal_entries_to_btree_roots(struct bch_fs *, struct jset *);
 313 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
 314                                         struct jset_entry *, struct jset_entry *);
 315
 316 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
 317 int bch2_fs_btree_interior_update_init(struct bch_fs *);
 318
 319 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */