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 int bch2_btree_split_leaf(struct bch_fs *, struct btree_iter *, unsigned);
 151
 152 int __bch2_foreground_maybe_merge(struct bch_fs *, struct btree_iter *,
 153                                   unsigned, enum btree_node_sibling);
 154
 155 static inline int bch2_foreground_maybe_merge_sibling(struct bch_fs *c,
 156                                         struct btree_iter *iter,
 157                                         unsigned level,
 158                                         enum btree_node_sibling sib)
 159 {
 160         struct btree *b;
 161
 162         if (!bch2_btree_node_relock(iter, level))
 163                 return 0;
 164
 165         b = iter->l[level].b;
 166         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
 167                 return 0;
 168
 169         return __bch2_foreground_maybe_merge(c, iter, level, sib);
 170 }
 171
 172 static inline void bch2_foreground_maybe_merge(struct bch_fs *c,
 173                                                struct btree_iter *iter,
 174                                                unsigned level)
 175 {
 176         bch2_foreground_maybe_merge_sibling(c, iter, level, btree_prev_sib);
 177         bch2_foreground_maybe_merge_sibling(c, iter, level, btree_next_sib);
 178 }
 179
 180 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
 181 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
 182
 183 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
 184                                                      struct btree *b)
 185 {
 186         unsigned depth = btree_node_root(c, b)->level - b->level;
 187
 188         return btree_reserve_required_nodes(depth);
 189 }
 190
 191 static inline void btree_node_reset_sib_u64s(struct btree *b)
 192 {
 193         b->sib_u64s[0] = b->nr.live_u64s;
 194         b->sib_u64s[1] = b->nr.live_u64s;
 195 }
 196
 197 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
 198 {
 199         return (void *) b->data + btree_bytes(c);
 200 }
 201
 202 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
 203                                                             struct btree *b)
 204 {
 205         return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
 206 }
 207
 208 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
 209                                                           struct btree *b)
 210 {
 211         return btree_data_end(c, b);
 212 }
 213
 214 static inline void *write_block(struct btree *b)
 215 {
 216         return (void *) b->data + (b->written << 9);
 217 }
 218
 219 static inline bool bset_written(struct btree *b, struct bset *i)
 220 {
 221         return (void *) i < write_block(b);
 222 }
 223
 224 static inline bool bset_unwritten(struct btree *b, struct bset *i)
 225 {
 226         return (void *) i > write_block(b);
 227 }
 228
 229 static inline unsigned bset_end_sector(struct bch_fs *c, struct btree *b,
 230                                        struct bset *i)
 231 {
 232         return round_up(bset_byte_offset(b, vstruct_end(i)),
 233                         block_bytes(c)) >> 9;
 234 }
 235
 236 static inline unsigned btree_write_set_buffer(struct btree *b)
 237 {
 238         /*
 239          * Could buffer up larger amounts of keys for btrees with larger keys,
 240          * pending benchmarking:
 241          */
 242         return 4 << 10;
 243 }
 244
 245 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 246                                                      struct btree *b)
 247 {
 248         struct bset *i = btree_bset_last(b);
 249         unsigned offset = max_t(unsigned, b->written << 9,
 250                                 bset_byte_offset(b, vstruct_end(i)));
 251         ssize_t remaining_space = (ssize_t) btree_bytes(c) - (ssize_t)
 252                 (offset + sizeof(struct btree_node_entry) +
 253                  b->whiteout_u64s * sizeof(u64) +
 254                  b->uncompacted_whiteout_u64s * sizeof(u64));
 255
 256         EBUG_ON(offset > btree_bytes(c));
 257
 258         if ((unlikely(bset_written(b, i)) &&
 259              remaining_space > block_bytes(c)) ||
 260             (unlikely(vstruct_bytes(i) > btree_write_set_buffer(b)) &&
 261              remaining_space > btree_write_set_buffer(b)))
 262                 return (void *) b->data + offset;
 263
 264         return NULL;
 265 }
 266
 267 static inline void unreserve_whiteout(struct btree *b, struct bset_tree *t,
 268                                       struct bkey_packed *k)
 269 {
 270         if (bset_written(b, bset(b, t))) {
 271                 EBUG_ON(b->uncompacted_whiteout_u64s <
 272                         bkeyp_key_u64s(&b->format, k));
 273                 b->uncompacted_whiteout_u64s -=
 274                         bkeyp_key_u64s(&b->format, k);
 275         }
 276 }
 277
 278 static inline void reserve_whiteout(struct btree *b, struct bset_tree *t,
 279                                     struct bkey_packed *k)
 280 {
 281         if (bset_written(b, bset(b, t))) {
 282                 BUG_ON(!k->needs_whiteout);
 283                 b->uncompacted_whiteout_u64s +=
 284                         bkeyp_key_u64s(&b->format, k);
 285         }
 286 }
 287
 288 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
 289                                                    struct btree *b)
 290 {
 291         struct bset *i = btree_bset_last(b);
 292         unsigned used = bset_byte_offset(b, vstruct_end(i)) / sizeof(u64) +
 293                 b->whiteout_u64s +
 294                 b->uncompacted_whiteout_u64s;
 295         unsigned total = c->opts.btree_node_size << 6;
 296
 297         EBUG_ON(used > total);
 298
 299         if (bset_written(b, i))
 300                 return 0;
 301
 302         return total - used;
 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_fake(b)))
 313                 return false;
 314
 315         if (btree_node_is_extents(b)) {
 316                 /* The insert key might split an existing key
 317                  * (bch2_insert_fixup_extent() -> BCH_EXTENT_OVERLAP_MIDDLE case:
 318                  */
 319                 u64s += BKEY_EXTENT_U64s_MAX;
 320         }
 321
 322         return u64s <= bch_btree_keys_u64s_remaining(c, b);
 323 }
 324
 325 static inline bool journal_res_insert_fits(struct btree_insert *trans,
 326                                            struct btree_insert_entry *insert)
 327 {
 328         unsigned u64s = 0;
 329         struct btree_insert_entry *i;
 330
 331         /*
 332          * If we didn't get a journal reservation, we're in journal replay and
 333          * we're not journalling updates:
 334          */
 335         if (!trans->journal_res.ref)
 336                 return true;
 337
 338         for (i = insert; i < trans->entries + trans->nr; i++)
 339                 u64s += jset_u64s(i->k->k.u64s + i->extra_res);
 340
 341         return u64s <= trans->journal_res.u64s;
 342 }
 343
 344 ssize_t bch2_btree_updates_print(struct bch_fs *, char *);
 345
 346 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *);
 347
 348 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */