#include "btree_gc.h"
#include "btree_io.h"
#include "btree_iter.h"
+#include "btree_key_cache.h"
#include "btree_locking.h"
#include "buckets.h"
#include "debug.h"
{
bch2_btree_node_lock_write(b, iter);
+ if (btree_iter_type(iter) == BTREE_ITER_CACHED)
+ return;
+
if (unlikely(btree_node_just_written(b)) &&
bch2_btree_post_write_cleanup(c, b))
bch2_btree_iter_reinit_node(iter, b);
EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
k = bch2_btree_node_iter_peek_all(node_iter, b);
- if (k && bkey_cmp_packed(b, k, &insert->k))
+ if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
k = NULL;
/* @k is the key being overwritten/deleted, if any: */
btree_node_lock_type(c, b, SIX_LOCK_read);
bch2_btree_node_write_cond(c, b,
(btree_current_write(b) == w && w->journal.seq == seq));
- six_unlock_read(&b->lock);
+ six_unlock_read(&b->c.lock);
}
static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
: btree_node_flush1);
}
-static inline void __btree_journal_key(struct btree_trans *trans,
- enum btree_id btree_id,
- struct bkey_i *insert)
-{
- struct journal *j = &trans->c->journal;
- u64 seq = trans->journal_res.seq;
- bool needs_whiteout = insert->k.needs_whiteout;
-
- /* ick */
- insert->k.needs_whiteout = false;
- bch2_journal_add_keys(j, &trans->journal_res,
- btree_id, insert);
- insert->k.needs_whiteout = needs_whiteout;
-
- bch2_journal_set_has_inode(j, &trans->journal_res,
- insert->k.p.inode);
-
- if (trans->journal_seq)
- *trans->journal_seq = seq;
-}
-
-static void bch2_btree_journal_key(struct btree_trans *trans,
- struct btree_iter *iter,
- struct bkey_i *insert)
-{
- struct bch_fs *c = trans->c;
- struct journal *j = &c->journal;
- struct btree *b = iter_l(iter)->b;
-
- EBUG_ON(trans->journal_res.ref !=
- !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
-
- if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
- __btree_journal_key(trans, iter->btree_id, insert);
- btree_bset_last(b)->journal_seq =
- cpu_to_le64(trans->journal_res.seq);
- }
-
- bch2_btree_add_journal_pin(c, b,
- likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
- ? trans->journal_res.seq
- : j->replay_journal_seq);
-
- if (unlikely(!btree_node_dirty(b)))
- set_btree_node_dirty(b);
-}
-
/**
* btree_insert_key - insert a key one key into a leaf node
*/
-static void btree_insert_key_leaf(struct btree_trans *trans,
+static bool btree_insert_key_leaf(struct btree_trans *trans,
struct btree_iter *iter,
struct bkey_i *insert)
{
struct bch_fs *c = trans->c;
struct btree *b = iter_l(iter)->b;
struct bset_tree *t = bset_tree_last(b);
+ struct bset *i = bset(b, t);
int old_u64s = bset_u64s(t);
int old_live_u64s = b->nr.live_u64s;
int live_u64s_added, u64s_added;
- insert->k.needs_whiteout = false;
+ EBUG_ON(!iter->level &&
+ !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
+
+ if (unlikely(!bch2_btree_bset_insert_key(iter, b,
+ &iter_l(iter)->iter, insert)))
+ return false;
- if (likely(bch2_btree_bset_insert_key(iter, b, &iter_l(iter)->iter, insert)))
- bch2_btree_journal_key(trans, iter, insert);
+ i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
+ le64_to_cpu(i->journal_seq)));
+
+ bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
+
+ if (unlikely(!btree_node_dirty(b)))
+ set_btree_node_dirty(c, b);
live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
u64s_added = (int) bset_u64s(t) - old_u64s;
bch2_btree_iter_reinit_node(iter, b);
trace_btree_insert_key(c, b, insert);
+ return true;
}
+/* Cached btree updates: */
+
/* Normal update interface: */
static inline void btree_insert_entry_checks(struct btree_trans *trans,
struct bch_fs *c = trans->c;
BUG_ON(bkey_cmp(insert->k.p, iter->pos));
- BUG_ON(debug_check_bkeys(c) &&
- bch2_bkey_invalid(c, bkey_i_to_s_c(insert), iter->btree_id));
+ BUG_ON(bch2_debug_check_bkeys &&
+ bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
+ __btree_node_type(iter->level, iter->btree_id)));
}
static noinline int
static enum btree_insert_ret
btree_key_can_insert(struct btree_trans *trans,
struct btree_iter *iter,
- struct bkey_i *insert,
- unsigned *u64s)
+ unsigned u64s)
{
struct bch_fs *c = trans->c;
struct btree *b = iter_l(iter)->b;
- static enum btree_insert_ret ret;
- if (unlikely(btree_node_fake(b)))
+ if (!bch2_btree_node_insert_fits(c, b, u64s))
return BTREE_INSERT_BTREE_NODE_FULL;
- /*
- * old bch2_extent_sort_fix_overlapping() algorithm won't work with new
- * style extent updates:
- */
- if (unlikely(btree_node_old_extent_overwrite(b)))
- return BTREE_INSERT_BTREE_NODE_FULL;
+ return BTREE_INSERT_OK;
+}
- ret = !btree_node_is_extents(b)
- ? BTREE_INSERT_OK
- : bch2_extent_can_insert(trans, iter, insert);
- if (ret)
- return ret;
+static enum btree_insert_ret
+btree_key_can_insert_cached(struct btree_trans *trans,
+ struct btree_iter *iter,
+ unsigned u64s)
+{
+ struct bkey_cached *ck = (void *) iter->l[0].b;
+ unsigned new_u64s;
+ struct bkey_i *new_k;
- if (*u64s > bch_btree_keys_u64s_remaining(c, b))
- return BTREE_INSERT_BTREE_NODE_FULL;
+ BUG_ON(iter->level);
+
+ if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
+ bch2_btree_key_cache_must_wait(trans->c))
+ return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
+
+ if (u64s <= ck->u64s)
+ return BTREE_INSERT_OK;
+
+ new_u64s = roundup_pow_of_two(u64s);
+ new_k = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
+ if (!new_k)
+ return -ENOMEM;
+ ck->u64s = new_u64s;
+ ck->k = new_k;
return BTREE_INSERT_OK;
}
struct btree_iter *iter,
struct bkey_i *insert)
{
- btree_insert_key_leaf(trans, iter, insert);
+ struct bch_fs *c = trans->c;
+ struct journal *j = &c->journal;
+ bool did_work;
+
+ EBUG_ON(trans->journal_res.ref !=
+ !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
+
+ insert->k.needs_whiteout = false;
+
+ did_work = (btree_iter_type(iter) != BTREE_ITER_CACHED)
+ ? btree_insert_key_leaf(trans, iter, insert)
+ : bch2_btree_insert_key_cached(trans, iter, insert);
+ if (!did_work)
+ return;
+
+ if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
+ bch2_journal_add_keys(j, &trans->journal_res,
+ iter->btree_id, insert);
+
+ bch2_journal_set_has_inode(j, &trans->journal_res,
+ insert->k.p.inode);
+
+ if (trans->journal_seq)
+ *trans->journal_seq = trans->journal_res.seq;
+ }
}
static inline bool iter_has_trans_triggers(struct btree_iter *iter)
static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
{
- return (BTREE_NODE_TYPE_HAS_TRIGGERS &
- ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) &
+ return (((BTREE_NODE_TYPE_HAS_TRIGGERS &
+ ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS)) |
+ (1U << BTREE_ID_EC)) &
(1U << iter->btree_id);
}
struct bch_fs *c = trans->c;
struct btree_insert_entry *i;
- trans_for_each_update(trans, i)
- if (gc_visited(c, gc_pos_btree_node(iter_l(i->iter)->b)))
+ trans_for_each_update(trans, i) {
+ /*
+ * XXX: synchronization of cached update triggers with gc
+ */
+ BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
+
+ if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
bch2_mark_update(trans, i->iter, i->k, NULL,
i->trigger_flags|BTREE_TRIGGER_GC);
+ }
}
static inline int
u64s = 0;
u64s += i->k->k.u64s;
- ret = btree_key_can_insert(trans, i->iter, i->k, &u64s);
+ ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
+ ? btree_key_can_insert(trans, i->iter, u64s)
+ : btree_key_can_insert_cached(trans, i->iter, u64s);
if (ret) {
*stopped_at = i;
return ret;
JOURNAL_RES_GET_NONBLOCK);
if (ret)
goto err;
+ } else {
+ trans->journal_res.seq = c->journal.replay_journal_seq;
+ }
+
+ if (unlikely(trans->extra_journal_entry_u64s)) {
+ memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
+ trans->extra_journal_entries,
+ trans->extra_journal_entry_u64s);
+
+ trans->journal_res.offset += trans->extra_journal_entry_u64s;
+ trans->journal_res.u64s -= trans->extra_journal_entry_u64s;
}
/*
*/
if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
- if (journal_seq_verify(c))
+ if (bch2_journal_seq_verify)
trans_for_each_update2(trans, i)
i->k->k.version.lo = trans->journal_res.seq;
- else if (inject_invalid_keys(c))
+ else if (bch2_inject_invalid_keys)
trans_for_each_update2(trans, i)
i->k->k.version = MAX_VERSION;
}
ret = bch2_journal_preres_get(&trans->c->journal,
&trans->journal_preres, trans->journal_preres_u64s,
- JOURNAL_RES_GET_NONBLOCK);
+ JOURNAL_RES_GET_NONBLOCK|
+ ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)
+ ? JOURNAL_RES_GET_RECLAIM : 0));
if (unlikely(ret == -EAGAIN))
ret = bch2_trans_journal_preres_get_cold(trans,
trans->journal_preres_u64s);
* or anything else that might call bch2_trans_relock(), since that
* would just retake the read locks:
*/
- trans_for_each_iter_all(trans, iter) {
+ trans_for_each_iter(trans, iter) {
if (iter->nodes_locked != iter->nodes_intent_locked) {
EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
EBUG_ON(trans->iters_live & (1ULL << iter->idx));
bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
i->iter);
+ if (!ret && trans->journal_pin)
+ bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
+ trans->journal_pin, NULL);
+
/*
* Drop journal reservation after dropping write locks, since dropping
* the journal reservation may kick off a journal write:
trans->nounlock = true;
trans_for_each_update2(trans, i)
- if (!same_leaf_as_prev(trans, i))
+ if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
+ !same_leaf_as_prev(trans, i))
bch2_foreground_maybe_merge(trans->c, i->iter,
0, trans->flags);
trans->nounlock = false;
- trans_for_each_update2(trans, i)
- bch2_btree_iter_downgrade(i->iter);
+ bch2_trans_downgrade(trans);
return 0;
}
trace_trans_restart_journal_res_get(trans->ip);
ret = -EINTR;
break;
- default:
- BUG_ON(ret >= 0);
- break;
- }
-
- if (ret == -EINTR) {
- int ret2 = bch2_btree_iter_traverse_all(trans);
+ case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
+ bch2_trans_unlock(trans);
- if (ret2) {
- trace_trans_restart_traverse(trans->ip);
- return ret2;
+ while (bch2_btree_key_cache_must_wait(c)) {
+ mutex_lock(&c->journal.reclaim_lock);
+ bch2_journal_reclaim(&c->journal);
+ mutex_unlock(&c->journal.reclaim_lock);
}
- trace_trans_restart_atomic(trans->ip);
+ if (bch2_trans_relock(trans))
+ return 0;
+
+ trace_trans_restart_journal_reclaim(trans->ip);
+ ret = -EINTR;
+ break;
+ default:
+ BUG_ON(ret >= 0);
+ break;
}
return ret;
return 0;
}
+static inline int btree_iter_pos_cmp(const struct btree_iter *l,
+ const struct btree_iter *r)
+{
+ return cmp_int(l->btree_id, r->btree_id) ?:
+ bkey_cmp(l->pos, r->pos);
+}
+
static void bch2_trans_update2(struct btree_trans *trans,
struct btree_iter *iter,
struct bkey_i *insert)
iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
trans_for_each_update2(trans, i) {
- if (btree_iter_cmp(n.iter, i->iter) == 0) {
+ if (btree_iter_pos_cmp(n.iter, i->iter) == 0) {
*i = n;
return;
}
- if (btree_iter_cmp(n.iter, i->iter) <= 0)
+ if (btree_iter_pos_cmp(n.iter, i->iter) <= 0)
break;
}
struct bkey_i *insert)
{
struct btree_iter *iter;
+ int ret;
+
+ ret = bch2_extent_can_insert(trans, orig_iter, insert);
+ if (ret)
+ return ret;
if (bkey_deleted(&insert->k))
return 0;
memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
- trans->journal_u64s = 0;
+ trans->journal_u64s = trans->extra_journal_entry_u64s;
trans->journal_preres_u64s = 0;
if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
return ret;
}
+#ifdef CONFIG_BCACHEFS_DEBUG
+ trans_for_each_update(trans, i)
+ if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
+ !(i->trigger_flags & BTREE_TRIGGER_NORUN))
+ bch2_btree_key_cache_verify_clean(trans,
+ i->iter->btree_id, i->iter->pos);
+#endif
+
/*
* Running triggers will append more updates to the list of updates as
* we're walking it:
trans_trigger_run = false;
trans_for_each_update(trans, i) {
- if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK)) {
+ if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK &&
+ (ret = bch2_btree_iter_traverse(i->iter)))) {
trace_trans_restart_traverse(trans->ip);
- ret = -EINTR;
goto out;
}
BUG_ON(i->iter->locks_want < 1);
u64s = jset_u64s(i->k->k.u64s);
- if (0)
+ if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
+ likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
trans->journal_preres_u64s += u64s;
trans->journal_u64s += u64s;
}
* Pending updates are kept sorted: first, find position of new update:
*/
trans_for_each_update(trans, i)
- if (btree_iter_cmp(iter, i->iter) <= 0)
+ if (btree_iter_pos_cmp(iter, i->iter) <= 0)
break;
/*
__bch2_btree_insert(&trans, id, k));
}
-int bch2_btree_delete_at_range(struct btree_trans *trans,
- struct btree_iter *iter,
- struct bpos end,
- u64 *journal_seq)
+int bch2_btree_delete_at(struct btree_trans *trans,
+ struct btree_iter *iter, unsigned flags)
+{
+ struct bkey_i k;
+
+ bkey_init(&k.k);
+ k.k.p = iter->pos;
+
+ bch2_trans_update(trans, iter, &k, 0);
+ return bch2_trans_commit(trans, NULL, NULL,
+ BTREE_INSERT_NOFAIL|
+ BTREE_INSERT_USE_RESERVE|flags);
+}
+
+int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
+ struct bpos start, struct bpos end,
+ u64 *journal_seq)
{
+ struct btree_iter *iter;
struct bkey_s_c k;
int ret = 0;
+
+ iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
+ ret = PTR_ERR_OR_ZERO(iter);
+ if (ret)
+ return ret;
retry:
while ((k = bch2_btree_iter_peek(iter)).k &&
!(ret = bkey_err(k)) &&
bkey_init(&delete.k);
+ /*
+ * This could probably be more efficient for extents:
+ */
+
/*
* For extents, iter.pos won't necessarily be the same as
* bkey_start_pos(k.k) (for non extents they always will be the
goto retry;
}
+ bch2_trans_iter_put(trans, iter);
return ret;
-
-}
-
-int bch2_btree_delete_at(struct btree_trans *trans,
- struct btree_iter *iter, unsigned flags)
-{
- struct bkey_i k;
-
- bkey_init(&k.k);
- k.k.p = iter->pos;
-
- bch2_trans_update(trans, iter, &k, 0);
- return bch2_trans_commit(trans, NULL, NULL,
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE|flags);
}
/*
struct bpos start, struct bpos end,
u64 *journal_seq)
{
- struct btree_trans trans;
- struct btree_iter *iter;
- int ret = 0;
-
- /*
- * XXX: whether we need mem/more iters depends on whether this btree id
- * has triggers
- */
- bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
-
- iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
-
- ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
- ret = bch2_trans_exit(&trans) ?: ret;
-
- BUG_ON(ret == -EINTR);
- return ret;
+ return bch2_trans_do(c, NULL, journal_seq, 0,
+ bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
}
projects / bcachefs-tools-debian / blobdiff