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 struct btree_reserve {
  10         struct disk_reservation disk_res;
  11         unsigned                nr;
  12         struct btree            *b[BTREE_RESERVE_MAX];
  13 };
  14
  15 void __bch2_btree_calc_format(struct bkey_format_state *, struct btree *);
  16 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *,
  17                                 struct bkey_format *);
  18
  19 /* Btree node freeing/allocation: */
  20
  21 /*
  22  * Tracks a btree node that has been (or is about to be) freed in memory, but
  23  * has _not_ yet been freed on disk (because the write that makes the new
  24  * node(s) visible and frees the old hasn't completed yet)
  25  */
  26 struct pending_btree_node_free {
  27         bool                    index_update_done;
  28
  29         __le64                  seq;
  30         enum btree_id           btree_id;
  31         unsigned                level;
  32         __BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
  33 };
  34
  35 /*
  36  * Tracks an in progress split/rewrite of a btree node and the update to the
  37  * parent node:
  38  *
  39  * When we split/rewrite a node, we do all the updates in memory without
  40  * waiting for any writes to complete - we allocate the new node(s) and update
  41  * the parent node, possibly recursively up to the root.
  42  *
  43  * The end result is that we have one or more new nodes being written -
  44  * possibly several, if there were multiple splits - and then a write (updating
  45  * an interior node) which will make all these new nodes visible.
  46  *
  47  * Additionally, as we split/rewrite nodes we free the old nodes - but the old
  48  * nodes can't be freed (their space on disk can't be reclaimed) until the
  49  * update to the interior node that makes the new node visible completes -
  50  * until then, the old nodes are still reachable on disk.
  51  *
  52  */
  53 struct btree_update {
  54         struct closure                  cl;
  55         struct bch_fs                   *c;
  56
  57         struct list_head                list;
  58         struct list_head                unwritten_list;
  59
  60         /* What kind of update are we doing? */
  61         enum {
  62                 BTREE_INTERIOR_NO_UPDATE,
  63                 BTREE_INTERIOR_UPDATING_NODE,
  64                 BTREE_INTERIOR_UPDATING_ROOT,
  65                 BTREE_INTERIOR_UPDATING_AS,
  66         } mode;
  67
  68         unsigned                        must_rewrite:1;
  69         unsigned                        nodes_written:1;
  70
  71         enum btree_id                   btree_id;
  72         u8                              level;
  73
  74         struct btree_reserve            *reserve;
  75         struct journal_preres           journal_preres;
  76
  77         /*
  78          * BTREE_INTERIOR_UPDATING_NODE:
  79          * The update that made the new nodes visible was a regular update to an
  80          * existing interior node - @b. We can't write out the update to @b
  81          * until the new nodes we created are finished writing, so we block @b
  82          * from writing by putting this btree_interior update on the
  83          * @b->write_blocked list with @write_blocked_list:
  84          */
  85         struct btree                    *b;
  86         struct list_head                write_blocked_list;
  87
  88         /*
  89          * We may be freeing nodes that were dirty, and thus had journal entries
  90          * pinned: we need to transfer the oldest of those pins to the
  91          * btree_update operation, and release it when the new node(s)
  92          * are all persistent and reachable:
  93          */
  94         struct journal_entry_pin        journal;
  95
  96         /*
  97          * Nodes being freed:
  98          * Protected by c->btree_node_pending_free_lock
  99          */
 100         struct pending_btree_node_free  pending[BTREE_MAX_DEPTH + GC_MERGE_NODES];
 101         unsigned                        nr_pending;
 102
 103         /* New nodes, that will be made reachable by this update: */
 104         struct btree                    *new_nodes[BTREE_MAX_DEPTH * 2 + GC_MERGE_NODES];
 105         unsigned                        nr_new_nodes;
 106
 107         unsigned                        journal_u64s;
 108         u64                             journal_entries[
 109                 (BKEY_BTREE_PTR_U64s_MAX + 1) * (BTREE_MAX_DEPTH - 1) * 2];
 110
 111         /* Only here to reduce stack usage on recursive splits: */
 112         struct keylist                  parent_keys;
 113         /*
 114          * Enough room for btree_split's keys without realloc - btree node
 115          * pointers never have crc/compression info, so we only need to acount
 116          * for the pointers for three keys
 117          */
 118         u64                             inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
 119 };
 120
 121 #define for_each_pending_btree_node_free(c, as, p)                      \
 122         list_for_each_entry(as, &c->btree_interior_update_list, list)   \
 123                 for (p = as->pending; p < as->pending + as->nr_pending; p++)
 124
 125 void bch2_btree_node_free_inmem(struct bch_fs *, struct btree *,
 126                                 struct btree_iter *);
 127 void bch2_btree_node_free_never_inserted(struct bch_fs *, struct btree *);
 128
 129 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
 130                                                   struct btree *,
 131                                                   struct bkey_format);
 132
 133 void bch2_btree_update_done(struct btree_update *);
 134 struct btree_update *
 135 bch2_btree_update_start(struct bch_fs *, enum btree_id, unsigned,
 136                         unsigned, struct closure *);
 137
 138 void bch2_btree_interior_update_will_free_node(struct btree_update *,
 139                                                struct btree *);
 140
 141 void bch2_btree_insert_node(struct btree_update *, struct btree *,
 142                             struct btree_iter *, struct keylist *,
 143                             unsigned);
 144 int bch2_btree_split_leaf(struct bch_fs *, struct btree_iter *, unsigned);
 145
 146 void __bch2_foreground_maybe_merge(struct bch_fs *, struct btree_iter *,
 147                                    unsigned, unsigned, enum btree_node_sibling);
 148
 149 static inline void bch2_foreground_maybe_merge_sibling(struct bch_fs *c,
 150                                         struct btree_iter *iter,
 151                                         unsigned level, unsigned flags,
 152                                         enum btree_node_sibling sib)
 153 {
 154         struct btree *b;
 155
 156         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 157                 return;
 158
 159         if (!bch2_btree_node_relock(iter, level))
 160                 return;
 161
 162         b = iter->l[level].b;
 163         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
 164                 return;
 165
 166         __bch2_foreground_maybe_merge(c, iter, level, flags, sib);
 167 }
 168
 169 static inline void bch2_foreground_maybe_merge(struct bch_fs *c,
 170                                                struct btree_iter *iter,
 171                                                unsigned level,
 172                                                unsigned flags)
 173 {
 174         bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
 175                                             btree_prev_sib);
 176         bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
 177                                             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 + 1;
 187
 188         /*
 189          * Number of nodes we might have to allocate in a worst case btree
 190          * split operation - we split all the way up to the root, then allocate
 191          * a new root, unless we're already at max depth:
 192          */
 193         if (depth < BTREE_MAX_DEPTH)
 194                 return (depth - b->level) * 2 + 1;
 195         else
 196                 return (depth - b->level) * 2 - 1;
 197 }
 198
 199 static inline void btree_node_reset_sib_u64s(struct btree *b)
 200 {
 201         b->sib_u64s[0] = b->nr.live_u64s;
 202         b->sib_u64s[1] = b->nr.live_u64s;
 203 }
 204
 205 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
 206 {
 207         return (void *) b->data + btree_bytes(c);
 208 }
 209
 210 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
 211                                                             struct btree *b)
 212 {
 213         return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
 214 }
 215
 216 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
 217                                                           struct btree *b)
 218 {
 219         return btree_data_end(c, b);
 220 }
 221
 222 static inline void *write_block(struct btree *b)
 223 {
 224         return (void *) b->data + (b->written << 9);
 225 }
 226
 227 static inline bool __btree_addr_written(struct btree *b, void *p)
 228 {
 229         return p < write_block(b);
 230 }
 231
 232 static inline bool bset_written(struct btree *b, struct bset *i)
 233 {
 234         return __btree_addr_written(b, i);
 235 }
 236
 237 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
 238 {
 239         return __btree_addr_written(b, k);
 240 }
 241
 242 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
 243                                                  struct btree *b,
 244                                                  void *end)
 245 {
 246         ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
 247                 b->whiteout_u64s;
 248         ssize_t total = c->opts.btree_node_size << 6;
 249
 250         return total - used;
 251 }
 252
 253 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
 254                                                    struct btree *b)
 255 {
 256         ssize_t remaining = __bch_btree_u64s_remaining(c, b,
 257                                 btree_bkey_last(b, bset_tree_last(b)));
 258
 259         BUG_ON(remaining < 0);
 260
 261         if (bset_written(b, btree_bset_last(b)))
 262                 return 0;
 263
 264         return remaining;
 265 }
 266
 267 static inline unsigned btree_write_set_buffer(struct btree *b)
 268 {
 269         /*
 270          * Could buffer up larger amounts of keys for btrees with larger keys,
 271          * pending benchmarking:
 272          */
 273         return 4 << 10;
 274 }
 275
 276 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 277                                                      struct btree *b)
 278 {
 279         struct bset_tree *t = bset_tree_last(b);
 280         struct btree_node_entry *bne = max(write_block(b),
 281                         (void *) btree_bkey_last(b, bset_tree_last(b)));
 282         ssize_t remaining_space =
 283                 __bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
 284
 285         if (unlikely(bset_written(b, bset(b, t)))) {
 286                 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
 287                         return bne;
 288         } else {
 289                 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
 290                     remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
 291                         return bne;
 292         }
 293
 294         return NULL;
 295 }
 296
 297 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
 298                                  struct bpos pos)
 299 {
 300         struct bkey_packed k;
 301
 302         BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
 303
 304         if (!bkey_pack_pos(&k, pos, b)) {
 305                 struct bkey *u = (void *) &k;
 306
 307                 bkey_init(u);
 308                 u->p = pos;
 309         }
 310
 311         k.needs_whiteout = true;
 312
 313         b->whiteout_u64s += k.u64s;
 314         bkey_copy(unwritten_whiteouts_start(c, b), &k);
 315 }
 316
 317 /*
 318  * write lock must be held on @b (else the dirty bset that we were going to
 319  * insert into could be written out from under us)
 320  */
 321 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
 322                                                struct btree *b, unsigned u64s)
 323 {
 324         if (unlikely(btree_node_fake(b)))
 325                 return false;
 326
 327         return u64s <= bch_btree_keys_u64s_remaining(c, b);
 328 }
 329
 330 ssize_t bch2_btree_updates_print(struct bch_fs *, char *);
 331
 332 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *);
 333
 334 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */