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
  73         struct btree_reserve            *reserve;
  74
  75         /*
  76          * BTREE_INTERIOR_UPDATING_NODE:
  77          * The update that made the new nodes visible was a regular update to an
  78          * existing interior node - @b. We can't write out the update to @b
  79          * until the new nodes we created are finished writing, so we block @b
  80          * from writing by putting this btree_interior update on the
  81          * @b->write_blocked list with @write_blocked_list:
  82          */
  83         struct btree                    *b;
  84         struct list_head                write_blocked_list;
  85
  86         /*
  87          * BTREE_INTERIOR_UPDATING_AS: btree node we updated was freed, so now
  88          * we're now blocking another btree_update
  89          * @parent_as - btree_update that's waiting on our nodes to finish
  90          * writing, before it can make new nodes visible on disk
  91          * @wait - list of child btree_updates that are waiting on this
  92          * btree_update to make all the new nodes visible before they can free
  93          * their old btree nodes
  94          */
  95         struct btree_update             *parent_as;
  96         struct closure_waitlist         wait;
  97
  98         /*
  99          * We may be freeing nodes that were dirty, and thus had journal entries
 100          * pinned: we need to transfer the oldest of those pins to the
 101          * btree_update operation, and release it when the new node(s)
 102          * are all persistent and reachable:
 103          */
 104         struct journal_entry_pin        journal;
 105
 106         u64                             journal_seq;
 107
 108         /*
 109          * Nodes being freed:
 110          * Protected by c->btree_node_pending_free_lock
 111          */
 112         struct pending_btree_node_free  pending[BTREE_MAX_DEPTH + GC_MERGE_NODES];
 113         unsigned                        nr_pending;
 114
 115         /* New nodes, that will be made reachable by this update: */
 116         struct btree                    *new_nodes[BTREE_MAX_DEPTH * 2 + GC_MERGE_NODES];
 117         unsigned                        nr_new_nodes;
 118
 119         /* Only here to reduce stack usage on recursive splits: */
 120         struct keylist                  parent_keys;
 121         /*
 122          * Enough room for btree_split's keys without realloc - btree node
 123          * pointers never have crc/compression info, so we only need to acount
 124          * for the pointers for three keys
 125          */
 126         u64                             inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
 127 };
 128
 129 #define for_each_pending_btree_node_free(c, as, p)                      \
 130         list_for_each_entry(as, &c->btree_interior_update_list, list)   \
 131                 for (p = as->pending; p < as->pending + as->nr_pending; p++)
 132
 133 void bch2_btree_node_free_inmem(struct bch_fs *, struct btree *,
 134                                 struct btree_iter *);
 135 void bch2_btree_node_free_never_inserted(struct bch_fs *, struct btree *);
 136
 137 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
 138                                                   struct btree *,
 139                                                   struct bkey_format);
 140
 141 void bch2_btree_update_done(struct btree_update *);
 142 struct btree_update *
 143 bch2_btree_update_start(struct bch_fs *, enum btree_id, unsigned,
 144                         unsigned, struct closure *);
 145
 146 void bch2_btree_interior_update_will_free_node(struct btree_update *,
 147                                                struct btree *);
 148
 149 void bch2_btree_insert_node(struct btree_update *, struct btree *,
 150                             struct btree_iter *, struct keylist *,
 151                             unsigned);
 152 int bch2_btree_split_leaf(struct bch_fs *, struct btree_iter *, unsigned);
 153
 154 void __bch2_foreground_maybe_merge(struct bch_fs *, struct btree_iter *,
 155                                    unsigned, unsigned, enum btree_node_sibling);
 156
 157 static inline void bch2_foreground_maybe_merge_sibling(struct bch_fs *c,
 158                                         struct btree_iter *iter,
 159                                         unsigned level, unsigned flags,
 160                                         enum btree_node_sibling sib)
 161 {
 162         struct btree *b;
 163
 164         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 165                 return;
 166
 167         if (!bch2_btree_node_relock(iter, level))
 168                 return;
 169
 170         b = iter->l[level].b;
 171         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
 172                 return;
 173
 174         __bch2_foreground_maybe_merge(c, iter, level, flags, sib);
 175 }
 176
 177 static inline void bch2_foreground_maybe_merge(struct bch_fs *c,
 178                                                struct btree_iter *iter,
 179                                                unsigned level,
 180                                                unsigned flags)
 181 {
 182         bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
 183                                             btree_prev_sib);
 184         bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
 185                                             btree_next_sib);
 186 }
 187
 188 void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
 189 void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
 190
 191 static inline unsigned btree_update_reserve_required(struct bch_fs *c,
 192                                                      struct btree *b)
 193 {
 194         unsigned depth = btree_node_root(c, b)->level + 1;
 195
 196         /*
 197          * Number of nodes we might have to allocate in a worst case btree
 198          * split operation - we split all the way up to the root, then allocate
 199          * a new root, unless we're already at max depth:
 200          */
 201         if (depth < BTREE_MAX_DEPTH)
 202                 return (depth - b->level) * 2 + 1;
 203         else
 204                 return (depth - b->level) * 2 - 1;
 205 }
 206
 207 static inline void btree_node_reset_sib_u64s(struct btree *b)
 208 {
 209         b->sib_u64s[0] = b->nr.live_u64s;
 210         b->sib_u64s[1] = b->nr.live_u64s;
 211 }
 212
 213 static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
 214 {
 215         return (void *) b->data + btree_bytes(c);
 216 }
 217
 218 static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
 219                                                             struct btree *b)
 220 {
 221         return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
 222 }
 223
 224 static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
 225                                                           struct btree *b)
 226 {
 227         return btree_data_end(c, b);
 228 }
 229
 230 static inline void *write_block(struct btree *b)
 231 {
 232         return (void *) b->data + (b->written << 9);
 233 }
 234
 235 static inline bool __btree_addr_written(struct btree *b, void *p)
 236 {
 237         return p < write_block(b);
 238 }
 239
 240 static inline bool bset_written(struct btree *b, struct bset *i)
 241 {
 242         return __btree_addr_written(b, i);
 243 }
 244
 245 static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
 246 {
 247         return __btree_addr_written(b, k);
 248 }
 249
 250 static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
 251                                                  struct btree *b,
 252                                                  void *end)
 253 {
 254         ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
 255                 b->whiteout_u64s;
 256         ssize_t total = c->opts.btree_node_size << 6;
 257
 258         return total - used;
 259 }
 260
 261 static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
 262                                                    struct btree *b)
 263 {
 264         ssize_t remaining = __bch_btree_u64s_remaining(c, b,
 265                                 btree_bkey_last(b, bset_tree_last(b)));
 266
 267         BUG_ON(remaining < 0);
 268
 269         if (bset_written(b, btree_bset_last(b)))
 270                 return 0;
 271
 272         return remaining;
 273 }
 274
 275 static inline unsigned btree_write_set_buffer(struct btree *b)
 276 {
 277         /*
 278          * Could buffer up larger amounts of keys for btrees with larger keys,
 279          * pending benchmarking:
 280          */
 281         return 4 << 10;
 282 }
 283
 284 static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 285                                                      struct btree *b)
 286 {
 287         struct bset_tree *t = bset_tree_last(b);
 288         struct btree_node_entry *bne = max(write_block(b),
 289                         (void *) btree_bkey_last(b, bset_tree_last(b)));
 290         ssize_t remaining_space =
 291                 __bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
 292
 293         if (unlikely(bset_written(b, bset(b, t)))) {
 294                 if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
 295                         return bne;
 296         } else {
 297                 if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
 298                     remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
 299                         return bne;
 300         }
 301
 302         return NULL;
 303 }
 304
 305 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
 306                                  struct bpos pos)
 307 {
 308         struct bkey_packed k;
 309
 310         BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
 311
 312         if (!bkey_pack_pos(&k, pos, b)) {
 313                 struct bkey *u = (void *) &k;
 314
 315                 bkey_init(u);
 316                 u->p = pos;
 317         }
 318
 319         k.needs_whiteout = true;
 320
 321         b->whiteout_u64s += k.u64s;
 322         bkey_copy(unwritten_whiteouts_start(c, b), &k);
 323 }
 324
 325 /*
 326  * write lock must be held on @b (else the dirty bset that we were going to
 327  * insert into could be written out from under us)
 328  */
 329 static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
 330                                                struct btree *b, unsigned u64s)
 331 {
 332         if (unlikely(btree_node_fake(b)))
 333                 return false;
 334
 335         return u64s <= bch_btree_keys_u64s_remaining(c, b);
 336 }
 337
 338 ssize_t bch2_btree_updates_print(struct bch_fs *, char *);
 339
 340 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *);
 341
 342 #endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */