2017-03-17 06:18:50 +00:00
|
|
|
/* SPDX-License-Identifier: GPL-2.0 */
|
|
|
|
#ifndef _BCACHEFS_BTREE_UPDATE_INTERIOR_H
|
|
|
|
#define _BCACHEFS_BTREE_UPDATE_INTERIOR_H
|
|
|
|
|
|
|
|
#include "btree_cache.h"
|
|
|
|
#include "btree_locking.h"
|
|
|
|
#include "btree_update.h"
|
|
|
|
|
|
|
|
void __bch2_btree_calc_format(struct bkey_format_state *, struct btree *);
|
|
|
|
bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *,
|
|
|
|
struct bkey_format *);
|
|
|
|
|
2020-05-25 18:57:06 +00:00
|
|
|
#define BTREE_UPDATE_NODES_MAX ((BTREE_MAX_DEPTH - 2) * 2 + GC_MERGE_NODES)
|
2017-03-17 06:18:50 +00:00
|
|
|
|
2020-05-25 18:57:06 +00:00
|
|
|
#define BTREE_UPDATE_JOURNAL_RES (BTREE_UPDATE_NODES_MAX * (BKEY_BTREE_PTR_U64s_MAX + 1))
|
2020-04-06 01:49:17 +00:00
|
|
|
|
2017-03-17 06:18:50 +00:00
|
|
|
/*
|
|
|
|
* Tracks an in progress split/rewrite of a btree node and the update to the
|
|
|
|
* parent node:
|
|
|
|
*
|
|
|
|
* When we split/rewrite a node, we do all the updates in memory without
|
|
|
|
* waiting for any writes to complete - we allocate the new node(s) and update
|
|
|
|
* the parent node, possibly recursively up to the root.
|
|
|
|
*
|
|
|
|
* The end result is that we have one or more new nodes being written -
|
|
|
|
* possibly several, if there were multiple splits - and then a write (updating
|
|
|
|
* an interior node) which will make all these new nodes visible.
|
|
|
|
*
|
|
|
|
* Additionally, as we split/rewrite nodes we free the old nodes - but the old
|
|
|
|
* nodes can't be freed (their space on disk can't be reclaimed) until the
|
|
|
|
* update to the interior node that makes the new node visible completes -
|
|
|
|
* until then, the old nodes are still reachable on disk.
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
struct btree_update {
|
|
|
|
struct closure cl;
|
|
|
|
struct bch_fs *c;
|
|
|
|
|
|
|
|
struct list_head list;
|
2020-02-08 21:39:37 +00:00
|
|
|
struct list_head unwritten_list;
|
2017-03-17 06:18:50 +00:00
|
|
|
|
|
|
|
/* What kind of update are we doing? */
|
|
|
|
enum {
|
|
|
|
BTREE_INTERIOR_NO_UPDATE,
|
|
|
|
BTREE_INTERIOR_UPDATING_NODE,
|
|
|
|
BTREE_INTERIOR_UPDATING_ROOT,
|
|
|
|
BTREE_INTERIOR_UPDATING_AS,
|
|
|
|
} mode;
|
|
|
|
|
|
|
|
unsigned nodes_written:1;
|
|
|
|
|
|
|
|
enum btree_id btree_id;
|
|
|
|
|
2020-05-25 18:57:06 +00:00
|
|
|
struct disk_reservation disk_res;
|
2020-02-09 00:06:31 +00:00
|
|
|
struct journal_preres journal_preres;
|
2017-03-17 06:18:50 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* BTREE_INTERIOR_UPDATING_NODE:
|
|
|
|
* The update that made the new nodes visible was a regular update to an
|
|
|
|
* existing interior node - @b. We can't write out the update to @b
|
|
|
|
* until the new nodes we created are finished writing, so we block @b
|
|
|
|
* from writing by putting this btree_interior update on the
|
|
|
|
* @b->write_blocked list with @write_blocked_list:
|
|
|
|
*/
|
|
|
|
struct btree *b;
|
|
|
|
struct list_head write_blocked_list;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We may be freeing nodes that were dirty, and thus had journal entries
|
|
|
|
* pinned: we need to transfer the oldest of those pins to the
|
|
|
|
* btree_update operation, and release it when the new node(s)
|
|
|
|
* are all persistent and reachable:
|
|
|
|
*/
|
|
|
|
struct journal_entry_pin journal;
|
|
|
|
|
2020-05-25 18:57:06 +00:00
|
|
|
/* Preallocated nodes we reserve when we start the update: */
|
|
|
|
struct btree *prealloc_nodes[BTREE_UPDATE_NODES_MAX];
|
|
|
|
unsigned nr_prealloc_nodes;
|
|
|
|
|
|
|
|
/* Nodes being freed: */
|
|
|
|
struct keylist old_keys;
|
|
|
|
u64 _old_keys[BTREE_UPDATE_NODES_MAX *
|
|
|
|
BKEY_BTREE_PTR_VAL_U64s_MAX];
|
|
|
|
|
|
|
|
/* Nodes being added: */
|
|
|
|
struct keylist new_keys;
|
|
|
|
u64 _new_keys[BTREE_UPDATE_NODES_MAX *
|
|
|
|
BKEY_BTREE_PTR_VAL_U64s_MAX];
|
2017-03-17 06:18:50 +00:00
|
|
|
|
|
|
|
/* New nodes, that will be made reachable by this update: */
|
2020-05-25 18:57:06 +00:00
|
|
|
struct btree *new_nodes[BTREE_UPDATE_NODES_MAX];
|
2017-03-17 06:18:50 +00:00
|
|
|
unsigned nr_new_nodes;
|
|
|
|
|
2020-06-09 19:44:03 +00:00
|
|
|
open_bucket_idx_t open_buckets[BTREE_UPDATE_NODES_MAX *
|
2020-05-25 18:57:06 +00:00
|
|
|
BCH_REPLICAS_MAX];
|
2020-06-09 19:44:03 +00:00
|
|
|
open_bucket_idx_t nr_open_buckets;
|
2020-05-25 18:57:06 +00:00
|
|
|
|
2020-03-31 20:23:43 +00:00
|
|
|
unsigned journal_u64s;
|
2020-04-06 01:49:17 +00:00
|
|
|
u64 journal_entries[BTREE_UPDATE_JOURNAL_RES];
|
2020-03-31 20:23:43 +00:00
|
|
|
|
2017-03-17 06:18:50 +00:00
|
|
|
/* Only here to reduce stack usage on recursive splits: */
|
|
|
|
struct keylist parent_keys;
|
|
|
|
/*
|
|
|
|
* Enough room for btree_split's keys without realloc - btree node
|
|
|
|
* pointers never have crc/compression info, so we only need to acount
|
|
|
|
* for the pointers for three keys
|
|
|
|
*/
|
|
|
|
u64 inline_keys[BKEY_BTREE_PTR_U64s_MAX * 3];
|
|
|
|
};
|
|
|
|
|
|
|
|
void bch2_btree_node_free_inmem(struct bch_fs *, struct btree *,
|
|
|
|
struct btree_iter *);
|
|
|
|
void bch2_btree_node_free_never_inserted(struct bch_fs *, struct btree *);
|
|
|
|
|
2020-05-25 18:57:06 +00:00
|
|
|
void bch2_btree_update_get_open_buckets(struct btree_update *, struct btree *);
|
|
|
|
|
2017-03-17 06:18:50 +00:00
|
|
|
struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
|
|
|
|
struct btree *,
|
|
|
|
struct bkey_format);
|
|
|
|
|
|
|
|
void bch2_btree_update_done(struct btree_update *);
|
|
|
|
struct btree_update *
|
2020-04-06 01:49:17 +00:00
|
|
|
bch2_btree_update_start(struct btree_trans *, enum btree_id, unsigned,
|
2017-03-17 06:18:50 +00:00
|
|
|
unsigned, struct closure *);
|
|
|
|
|
|
|
|
void bch2_btree_interior_update_will_free_node(struct btree_update *,
|
|
|
|
struct btree *);
|
2020-05-25 18:57:06 +00:00
|
|
|
void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
|
2017-03-17 06:18:50 +00:00
|
|
|
|
|
|
|
void bch2_btree_insert_node(struct btree_update *, struct btree *,
|
|
|
|
struct btree_iter *, struct keylist *,
|
|
|
|
unsigned);
|
|
|
|
int bch2_btree_split_leaf(struct bch_fs *, struct btree_iter *, unsigned);
|
|
|
|
|
|
|
|
void __bch2_foreground_maybe_merge(struct bch_fs *, struct btree_iter *,
|
|
|
|
unsigned, unsigned, enum btree_node_sibling);
|
|
|
|
|
|
|
|
static inline void bch2_foreground_maybe_merge_sibling(struct bch_fs *c,
|
|
|
|
struct btree_iter *iter,
|
|
|
|
unsigned level, unsigned flags,
|
|
|
|
enum btree_node_sibling sib)
|
|
|
|
{
|
|
|
|
struct btree *b;
|
|
|
|
|
|
|
|
if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (!bch2_btree_node_relock(iter, level))
|
|
|
|
return;
|
|
|
|
|
|
|
|
b = iter->l[level].b;
|
|
|
|
if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
|
|
|
|
return;
|
|
|
|
|
|
|
|
__bch2_foreground_maybe_merge(c, iter, level, flags, sib);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void bch2_foreground_maybe_merge(struct bch_fs *c,
|
|
|
|
struct btree_iter *iter,
|
|
|
|
unsigned level,
|
|
|
|
unsigned flags)
|
|
|
|
{
|
|
|
|
bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
|
|
|
|
btree_prev_sib);
|
|
|
|
bch2_foreground_maybe_merge_sibling(c, iter, level, flags,
|
|
|
|
btree_next_sib);
|
|
|
|
}
|
|
|
|
|
|
|
|
void bch2_btree_set_root_for_read(struct bch_fs *, struct btree *);
|
|
|
|
void bch2_btree_root_alloc(struct bch_fs *, enum btree_id);
|
|
|
|
|
|
|
|
static inline unsigned btree_update_reserve_required(struct bch_fs *c,
|
|
|
|
struct btree *b)
|
|
|
|
{
|
2020-06-06 16:28:01 +00:00
|
|
|
unsigned depth = btree_node_root(c, b)->c.level + 1;
|
2017-03-17 06:18:50 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Number of nodes we might have to allocate in a worst case btree
|
|
|
|
* split operation - we split all the way up to the root, then allocate
|
|
|
|
* a new root, unless we're already at max depth:
|
|
|
|
*/
|
|
|
|
if (depth < BTREE_MAX_DEPTH)
|
2020-06-06 16:28:01 +00:00
|
|
|
return (depth - b->c.level) * 2 + 1;
|
2017-03-17 06:18:50 +00:00
|
|
|
else
|
2020-06-06 16:28:01 +00:00
|
|
|
return (depth - b->c.level) * 2 - 1;
|
2017-03-17 06:18:50 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline void btree_node_reset_sib_u64s(struct btree *b)
|
|
|
|
{
|
|
|
|
b->sib_u64s[0] = b->nr.live_u64s;
|
|
|
|
b->sib_u64s[1] = b->nr.live_u64s;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void *btree_data_end(struct bch_fs *c, struct btree *b)
|
|
|
|
{
|
|
|
|
return (void *) b->data + btree_bytes(c);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct bkey_packed *unwritten_whiteouts_start(struct bch_fs *c,
|
|
|
|
struct btree *b)
|
|
|
|
{
|
|
|
|
return (void *) ((u64 *) btree_data_end(c, b) - b->whiteout_u64s);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct bkey_packed *unwritten_whiteouts_end(struct bch_fs *c,
|
|
|
|
struct btree *b)
|
|
|
|
{
|
|
|
|
return btree_data_end(c, b);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void *write_block(struct btree *b)
|
|
|
|
{
|
|
|
|
return (void *) b->data + (b->written << 9);
|
|
|
|
}
|
|
|
|
|
2018-08-06 02:23:44 +00:00
|
|
|
static inline bool __btree_addr_written(struct btree *b, void *p)
|
|
|
|
{
|
|
|
|
return p < write_block(b);
|
|
|
|
}
|
|
|
|
|
2017-03-17 06:18:50 +00:00
|
|
|
static inline bool bset_written(struct btree *b, struct bset *i)
|
|
|
|
{
|
2018-08-06 02:23:44 +00:00
|
|
|
return __btree_addr_written(b, i);
|
2017-03-17 06:18:50 +00:00
|
|
|
}
|
|
|
|
|
2018-08-06 02:23:44 +00:00
|
|
|
static inline bool bkey_written(struct btree *b, struct bkey_packed *k)
|
2017-03-17 06:18:50 +00:00
|
|
|
{
|
2018-08-06 02:23:44 +00:00
|
|
|
return __btree_addr_written(b, k);
|
2017-03-17 06:18:50 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
|
|
|
|
struct btree *b,
|
|
|
|
void *end)
|
|
|
|
{
|
|
|
|
ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
|
2019-11-29 19:08:51 +00:00
|
|
|
b->whiteout_u64s;
|
2017-03-17 06:18:50 +00:00
|
|
|
ssize_t total = c->opts.btree_node_size << 6;
|
|
|
|
|
2020-11-11 17:42:54 +00:00
|
|
|
/* Always leave one extra u64 for bch2_varint_decode: */
|
|
|
|
used++;
|
|
|
|
|
2017-03-17 06:18:50 +00:00
|
|
|
return total - used;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline size_t bch_btree_keys_u64s_remaining(struct bch_fs *c,
|
|
|
|
struct btree *b)
|
|
|
|
{
|
|
|
|
ssize_t remaining = __bch_btree_u64s_remaining(c, b,
|
|
|
|
btree_bkey_last(b, bset_tree_last(b)));
|
|
|
|
|
|
|
|
BUG_ON(remaining < 0);
|
|
|
|
|
|
|
|
if (bset_written(b, btree_bset_last(b)))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
return remaining;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline unsigned btree_write_set_buffer(struct btree *b)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
* Could buffer up larger amounts of keys for btrees with larger keys,
|
|
|
|
* pending benchmarking:
|
|
|
|
*/
|
|
|
|
return 4 << 10;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
|
|
|
|
struct btree *b)
|
|
|
|
{
|
2019-10-28 23:35:13 +00:00
|
|
|
struct bset_tree *t = bset_tree_last(b);
|
2017-03-17 06:18:50 +00:00
|
|
|
struct btree_node_entry *bne = max(write_block(b),
|
|
|
|
(void *) btree_bkey_last(b, bset_tree_last(b)));
|
|
|
|
ssize_t remaining_space =
|
|
|
|
__bch_btree_u64s_remaining(c, b, &bne->keys.start[0]);
|
|
|
|
|
2019-10-28 23:35:13 +00:00
|
|
|
if (unlikely(bset_written(b, bset(b, t)))) {
|
2017-03-17 06:18:50 +00:00
|
|
|
if (remaining_space > (ssize_t) (block_bytes(c) >> 3))
|
|
|
|
return bne;
|
|
|
|
} else {
|
2019-10-28 23:35:13 +00:00
|
|
|
if (unlikely(bset_u64s(t) * sizeof(u64) > btree_write_set_buffer(b)) &&
|
2017-03-17 06:18:50 +00:00
|
|
|
remaining_space > (ssize_t) (btree_write_set_buffer(b) >> 3))
|
|
|
|
return bne;
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2019-11-29 19:08:51 +00:00
|
|
|
static inline void push_whiteout(struct bch_fs *c, struct btree *b,
|
2019-12-30 19:37:25 +00:00
|
|
|
struct bpos pos)
|
2017-03-17 06:18:50 +00:00
|
|
|
{
|
2019-12-30 19:37:25 +00:00
|
|
|
struct bkey_packed k;
|
2017-03-17 06:18:50 +00:00
|
|
|
|
2019-12-30 19:37:25 +00:00
|
|
|
BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
|
2019-11-29 19:08:51 +00:00
|
|
|
|
2019-12-30 19:37:25 +00:00
|
|
|
if (!bkey_pack_pos(&k, pos, b)) {
|
|
|
|
struct bkey *u = (void *) &k;
|
|
|
|
|
|
|
|
bkey_init(u);
|
|
|
|
u->p = pos;
|
|
|
|
}
|
|
|
|
|
|
|
|
k.needs_whiteout = true;
|
|
|
|
|
|
|
|
b->whiteout_u64s += k.u64s;
|
|
|
|
bkey_copy(unwritten_whiteouts_start(c, b), &k);
|
2017-03-17 06:18:50 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* write lock must be held on @b (else the dirty bset that we were going to
|
|
|
|
* insert into could be written out from under us)
|
|
|
|
*/
|
|
|
|
static inline bool bch2_btree_node_insert_fits(struct bch_fs *c,
|
2018-08-05 19:21:52 +00:00
|
|
|
struct btree *b, unsigned u64s)
|
2017-03-17 06:18:50 +00:00
|
|
|
{
|
2020-07-25 18:19:37 +00:00
|
|
|
if (unlikely(btree_node_need_rewrite(b)))
|
2017-03-17 06:18:50 +00:00
|
|
|
return false;
|
|
|
|
|
|
|
|
return u64s <= bch_btree_keys_u64s_remaining(c, b);
|
|
|
|
}
|
|
|
|
|
2020-07-25 21:06:11 +00:00
|
|
|
void bch2_btree_updates_to_text(struct printbuf *, struct bch_fs *);
|
2017-03-17 06:18:50 +00:00
|
|
|
|
|
|
|
size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *);
|
|
|
|
|
2020-05-25 18:57:06 +00:00
|
|
|
void bch2_journal_entries_to_btree_roots(struct bch_fs *, struct jset *);
|
|
|
|
struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
|
|
|
|
struct jset_entry *, struct jset_entry *);
|
|
|
|
|
2020-05-26 00:35:53 +00:00
|
|
|
void bch2_fs_btree_interior_update_exit(struct bch_fs *);
|
|
|
|
int bch2_fs_btree_interior_update_init(struct bch_fs *);
|
|
|
|
|
2017-03-17 06:18:50 +00:00
|
|
|
#endif /* _BCACHEFS_BTREE_UPDATE_INTERIOR_H */
|