git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_interior.h

   1 #ifndef _BCACHEFS_BTREE_UPDATE_INTERIOR_H
   2 #define _BCACHEFS_BTREE_UPDATE_INTERIOR_H
   3
   4 #include "btree_cache.h"
   5 #include "btree_locking.h"
   6 #include "btree_update.h"
   7
   8 struct btree_reserve {
   9         struct disk_reservation disk_res;
  10         unsigned                nr;
  11         struct btree            *b[BTREE_RESERVE_MAX];
  12 };
  13
  14 void __bch2_btree_calc_format(struct bkey_format_state *, struct btree *);
  15 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *,
  16                                 struct bkey_format *);
  17
  18 /* Btree node freeing/allocation: */
  19
  20 /*
  21  * Tracks a btree node that has been (or is about to be) freed in memory, but
  22  * has _not_ yet been freed on disk (because the write that makes the new
  23  * node(s) visible and frees the old hasn't completed yet)
  24  */
  25 struct pending_btree_node_free {
  26         bool                    index_update_done;
  27
  28         __le64                  seq;
  29         enum btree_id           btree_id;
  30         unsigned                level;
  31         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
  32 };
  33
  34 /*
  35  * Tracks an in progress split/rewrite of a btree node and the update to the
  36  * parent node:
  37  *
  38  * When we split/rewrite a node, we do all the updates in memory without
  39  * waiting for any writes to complete - we allocate the new node(s) and update
  40  * the parent node, possibly recursively up to the root.
  41  *
  42  * The end result is that we have one or more new nodes being written -
  43  * possibly several, if there were multiple splits - and then a write (updating
  44  * an interior node) which will make all these new nodes visible.
  45  *
  46  * Additionally, as we split/rewrite nodes we free the old nodes - but the old
  47  * nodes can't be freed (their space on disk can't be reclaimed) until the
  48  * update to the interior node that makes the new node visible completes -
  49  * until then, the old nodes are still reachable on disk.
  50  *
  51  */
  52 struct btree_update {
  53         struct closure                  cl;
  54         struct bch_fs                   *c;
  55
  56         struct list_head                list;
  57
  58         /* What kind of update are we doing? */
  59         enum {
  60                 BTREE_INTERIOR_NO_UPDATE,
  61                 BTREE_INTERIOR_UPDATING_NODE,
  62                 BTREE_INTERIOR_UPDATING_ROOT,
  63                 BTREE_INTERIOR_UPDATING_AS,
  64         } mode;
  65
  66         unsigned                        must_rewrite:1;
  67         unsigned                        nodes_written:1;
  68
  69         enum btree_id                   btree_id;
  70
  71         struct btree_reserve            *reserve;
  72
  73         /*
  74          * BTREE_INTERIOR_UPDATING_NODE:
  75          * The update that made the new nodes visible was a regular update to an
  76          * existing interior node - @b. We can't write out the update to @b
  77          * until the new nodes we created are finished writing, so we block @b
  78          * from writing by putting this btree_interior update on the
  79          * @b->write_blocked list with @write_blocked_list:
  80          */
  81         struct btree                    *b;
  82         struct list_head                write_blocked_list;
  83
  84         /*
  85          * BTREE_INTERIOR_UPDATING_AS: btree node we updated was freed, so now
  86          * we're now blocking another btree_update
  87          * @parent_as - btree_update that's waiting on our nodes to finish
  88          * writing, before it can make new nodes visible on disk
  89          * @wait - list of child btree_updates that are waiting on this
  90          * btree_update to make all the new nodes visible before they can free
  91          * their old btree nodes
  92          */
  93         struct btree_update             *parent_as;
  94         struct closure_waitlist         wait;
  95
  96         /*
  97          * We may be freeing nodes that were dirty, and thus had journal entries
  98          * pinned: we need to transfer the oldest of those pins to the
  99          * btree_update operation, and release it when the new node(s)
 100          * are all persistent and reachable:
 101          */
 102         struct journal_entry_pin        journal;
 103
 104         u64                             journal_seq;
 105
 106         /*
 107          * Nodes being freed:
 108          * Protected by c->btree_node_pending_free_lock
 109          */
 110         struct pending_btree_node_free  pending[BTREE_MAX_DEPTH + GC_MERGE_NODES];
 111         unsigned                        nr_pending;
 112
 113         /* New nodes, that will be made reachable by this update: */
 114         struct btree                    *new_nodes[BTREE_MAX_DEPTH * 2 + GC_MERGE_NODES];
 115         unsigned                        nr_new_nodes;
 116
 117         /* Only here to reduce stack usage on recursive splits: */
 118         struct keylist                  parent_keys;
 119         /*
 120          * Enough room for btree_split's keys without realloc - btree node
 121          * pointers never have crc/compression info, so we only need to acount
 122          * for the pointers for three keys
 123          */
 124         u64                             inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
 125 };
 126
 127 #define for_each_pending_btree_node_free(c, as, p)                      \
 128         list_for_each_entry(as, &c->btree_interior_update_list, list)   \
 129                 for (p = as->pending; p < as->pending + as->nr_pending; p++)
 130
 131 void bch2_btree_node_free_inmem(struct bch_fs *, struct btree *,
 132                                 struct btree_iter *);
 133 void bch2_btree_node_free_never_inserted(struct bch_fs *, struct btree *);
 134 void bch2_btree_open_bucket_put(struct bch_fs *, struct btree *);
 135
 136 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
 137                                                   struct btree *,
 138                                                   struct bkey_format);
 139
 140 void bch2_btree_update_done(struct btree_update *);
 141 struct btree_update *
 142 bch2_btree_update_start(struct bch_fs *, enum btree_id, unsigned,
 143                         unsigned, struct closure *);
 144
 145 void bch2_btree_interior_update_will_free_node(struct btree_update *,
 146                                                struct btree *);
 147
 148 void bch2_btree_insert_node(struct btree_update *, struct btree *,
 149                             struct btree_iter *, struct keylist *,
 150                             unsigned);
 151 int bch2_btree_split_leaf(struct bch_fs *, struct btree_iter *, unsigned);
 152
 153 void __bch2_foreground_maybe_merge(struct bch_fs *, struct btree_iter *,
 154                                    unsigned, unsigned, enum btree_node_sibling);
 155
 156 static inline void bch2_foreground_maybe_merge_sibling(struct bch_fs *c,
 157                                         struct btree_iter *iter,
 158                                         unsigned level, unsigned flags,
 159                                         enum btree_node_sibling sib)
 160 {
 161         struct btree *b;
 162
 163         /*
 164          * iterators are inconsistent when they hit end of leaf, until
 165          * traversed again
 166          *
 167          * XXX inconsistent how?
 168          */
 169         if (iter->flags & BTREE_ITER_AT_END_OF_LEAF)
 170                 return;
 171
 172         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 173                 return;
 174
 175         if (!bch2_btree_node_relock(iter, level))
 176                 return;
 177
 178         b = iter->l[level].b;
 179         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
 180                 return;
 181
 182         __bch2_foreground_maybe_merge(c, iter, level, flags, sib);
 183 }
 184
 185 static inline void bch2_foreground_maybe_merge(struct bch_fs *c,
 186                                                struct btree_iter *iter,
 187                                                unsigned level,
 188                                                unsigned flags)
 189 {
 190         bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
 191                                             btree_prev_sib);
 192         bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
 193                                             btree_next_sib);
 194 }
 195
 196 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
 197 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
 198
 199 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
 200                                                      struct btree *b)
 201 {
 202         unsigned depth = btree_node_root(c, b)->level + 1;
 203
 204         /*
 205          * Number of nodes we might have to allocate in a worst case btree
 206          * split operation - we split all the way up to the root, then allocate
 207          * a new root, unless we're already at max depth:
 208          */
 209         if (depth < BTREE_MAX_DEPTH)
 210                 return (depth - b->level) * 2 + 1;
 211         else
 212                 return (depth - b->level) * 2 - 1;
 213 }
 214
 215 static inline void btree_node_reset_sib_u64s(struct btree *b)
 216 {
 217         b->sib_u64s[0] = b->nr.live_u64s;
 218         b->sib_u64s[1] = b->nr.live_u64s;
 219 }
 220
 221 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
 222 {
 223         return (void *) b->data + btree_bytes(c);
 224 }
 225
 226 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
 227                                                             struct btree *b)
 228 {
 229         return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
 230 }
 231
 232 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
 233                                                           struct btree *b)
 234 {
 235         return btree_data_end(c, b);
 236 }
 237
 238 static inline void *write_block(struct btree *b)
 239 {
 240         return (void *) b->data + (b->written << 9);
 241 }
 242
 243 static inline bool bset_written(struct btree *b, struct bset *i)
 244 {
 245         return (void *) i < write_block(b);
 246 }
 247
 248 static inline bool bset_unwritten(struct btree *b, struct bset *i)
 249 {
 250         return (void *) i > write_block(b);
 251 }
 252
 253 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
 254                                                  struct btree *b,
 255                                                  void *end)
 256 {
 257         ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
 258                 b->whiteout_u64s +
 259                 b->uncompacted_whiteout_u64s;
 260         ssize_t total = c->opts.btree_node_size << 6;
 261
 262         return total - used;
 263 }
 264
 265 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
 266                                                    struct btree *b)
 267 {
 268         ssize_t remaining = __bch_btree_u64s_remaining(c, b,
 269                                 btree_bkey_last(b, bset_tree_last(b)));
 270
 271         BUG_ON(remaining < 0);
 272
 273         if (bset_written(b, btree_bset_last(b)))
 274                 return 0;
 275
 276         return remaining;
 277 }
 278
 279 static inline unsigned btree_write_set_buffer(struct btree *b)
 280 {
 281         /*
 282          * Could buffer up larger amounts of keys for btrees with larger keys,
 283          * pending benchmarking:
 284          */
 285         return 4 << 10;
 286 }
 287
 288 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 289                                                      struct btree *b)
 290 {
 291         struct bset *i = btree_bset_last(b);
 292         struct btree_node_entry *bne = max(write_block(b),
 293                         (void *) btree_bkey_last(b, bset_tree_last(b)));
 294         ssize_t remaining_space =
 295                 __bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
 296
 297         if (unlikely(bset_written(b, i))) {
 298                 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
 299                         return bne;
 300         } else {
 301                 if (unlikely(vstruct_bytes(i) > btree_write_set_buffer(b)) &&
 302                     remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
 303                         return bne;
 304         }
 305
 306         return NULL;
 307 }
 308
 309 static inline void unreserve_whiteout(struct btree *b, struct bset_tree *t,
 310                                       struct bkey_packed *k)
 311 {
 312         if (bset_written(b, bset(b, t))) {
 313                 EBUG_ON(b->uncompacted_whiteout_u64s <
 314                         bkeyp_key_u64s(&b->format, k));
 315                 b->uncompacted_whiteout_u64s -=
 316                         bkeyp_key_u64s(&b->format, k);
 317         }
 318 }
 319
 320 static inline void reserve_whiteout(struct btree *b, struct bset_tree *t,
 321                                     struct bkey_packed *k)
 322 {
 323         if (bset_written(b, bset(b, t))) {
 324                 BUG_ON(!k->needs_whiteout);
 325                 b->uncompacted_whiteout_u64s +=
 326                         bkeyp_key_u64s(&b->format, k);
 327         }
 328 }
 329
 330 /*
 331  * write lock must be held on @b (else the dirty bset that we were going to
 332  * insert into could be written out from under us)
 333  */
 334 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
 335                                               struct btree *b, unsigned u64s)
 336 {
 337         if (unlikely(btree_node_fake(b)))
 338                 return false;
 339
 340         if (btree_node_is_extents(b)) {
 341                 /* The insert key might split an existing key
 342                  * (bch2_insert_fixup_extent() -> BCH_EXTENT_OVERLAP_MIDDLE case:
 343                  */
 344                 u64s += BKEY_EXTENT_U64s_MAX;
 345         }
 346
 347         return u64s <= bch_btree_keys_u64s_remaining(c, b);
 348 }
 349
 350 static inline bool journal_res_insert_fits(struct btree_insert *trans,
 351                                            struct btree_insert_entry *insert)
 352 {
 353         unsigned u64s = 0;
 354         struct btree_insert_entry *i;
 355
 356         /*
 357          * If we didn't get a journal reservation, we're in journal replay and
 358          * we're not journalling updates:
 359          */
 360         if (!trans->journal_res.ref)
 361                 return true;
 362
 363         for (i = insert; i < trans->entries + trans->nr; i++)
 364                 u64s += jset_u64s(i->k->k.u64s + i->extra_res);
 365
 366         return u64s <= trans->journal_res.u64s;
 367 }
 368
 369 ssize_t bch2_btree_updates_print(struct bch_fs *, char *);
 370
 371 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *);
 372
 373 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */