#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"
#include <trace/events/bcachefs.h>
static inline bool same_leaf_as_prev(struct btree_trans *trans,
- unsigned idx)
+ struct btree_insert_entry *i)
{
- return idx &&
- trans->updates[trans->updates_sorted[idx]].iter->l[0].b ==
- trans->updates[trans->updates_sorted[idx - 1]].iter->l[0].b;
+ return i != trans->updates2 &&
+ iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
}
-#define trans_for_each_update_sorted(_trans, _i, _iter) \
- for (_iter = 0; \
- _iter < _trans->nr_updates && \
- (_i = _trans->updates + _trans->updates_sorted[_iter], 1); \
- _iter++)
-
inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
struct btree_iter *iter)
{
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);
bch2_btree_init_next(c, b, iter);
}
-static inline void btree_trans_sort_updates(struct btree_trans *trans)
-{
- struct btree_insert_entry *l, *r;
- unsigned nr = 0, pos;
-
- trans_for_each_update(trans, l) {
- for (pos = 0; pos < nr; pos++) {
- r = trans->updates + trans->updates_sorted[pos];
-
- if (btree_iter_cmp(l->iter, r->iter) <= 0)
- break;
- }
-
- memmove(&trans->updates_sorted[pos + 1],
- &trans->updates_sorted[pos],
- (nr - pos) * sizeof(trans->updates_sorted[0]));
-
- trans->updates_sorted[pos] = l - trans->updates;
- nr++;
- }
-}
-
/* Inserting into a given leaf node (last stage of insert): */
/* Handle overwrites and do insert, for non extents: */
struct btree_node_iter *node_iter,
struct bkey_i *insert)
{
- const struct bkey_format *f = &b->format;
struct bkey_packed *k;
- unsigned clobber_u64s;
+ unsigned clobber_u64s = 0, new_u64s = 0;
EBUG_ON(btree_node_just_written(b));
EBUG_ON(bset_written(b, btree_bset_last(b)));
EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
- EBUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0 ||
- bkey_cmp(insert->k.p, b->data->max_key) > 0);
+ EBUG_ON(bkey_cmp(b->data->min_key, POS_MIN) &&
+ bkey_cmp(bkey_start_pos(&insert->k),
+ bkey_predecessor(b->data->min_key)) < 0);
+ EBUG_ON(bkey_cmp(insert->k.p, b->data->min_key) < 0);
+ EBUG_ON(bkey_cmp(insert->k.p, b->data->max_key) > 0);
+ EBUG_ON(insert->k.u64s >
+ bch_btree_keys_u64s_remaining(iter->trans->c, 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)) {
- BUG_ON(bkey_whiteout(k));
-
- if (!bkey_written(b, k) &&
- bkey_val_u64s(&insert->k) == bkeyp_val_u64s(f, k) &&
- !bkey_whiteout(&insert->k)) {
- k->type = insert->k.type;
- memcpy_u64s(bkeyp_val(f, k), &insert->v,
- bkey_val_u64s(&insert->k));
- return true;
- }
+ if (k && bkey_cmp_packed(b, k, &insert->k))
+ k = NULL;
- insert->k.needs_whiteout = k->needs_whiteout;
+ /* @k is the key being overwritten/deleted, if any: */
+ EBUG_ON(k && bkey_whiteout(k));
+
+ /* Deleting, but not found? nothing to do: */
+ if (bkey_whiteout(&insert->k) && !k)
+ return false;
+ if (bkey_whiteout(&insert->k)) {
+ /* Deleting: */
btree_account_key_drop(b, k);
+ k->type = KEY_TYPE_deleted;
+
+ if (k->needs_whiteout)
+ push_whiteout(iter->trans->c, b, insert->k.p);
+ k->needs_whiteout = false;
if (k >= btree_bset_last(b)->start) {
clobber_u64s = k->u64s;
-
- /*
- * If we're deleting, and the key we're deleting doesn't
- * need a whiteout (it wasn't overwriting a key that had
- * been written to disk) - just delete it:
- */
- if (bkey_whiteout(&insert->k) && !k->needs_whiteout) {
- bch2_bset_delete(b, k, clobber_u64s);
- bch2_btree_node_iter_fix(iter, b, node_iter,
- k, clobber_u64s, 0);
- return true;
- }
-
- goto overwrite;
+ bch2_bset_delete(b, k, clobber_u64s);
+ goto fix_iter;
+ } else {
+ bch2_btree_iter_fix_key_modified(iter, b, k);
}
+ return true;
+ }
+
+ if (k) {
+ /* Overwriting: */
+ btree_account_key_drop(b, k);
k->type = KEY_TYPE_deleted;
- bch2_btree_node_iter_fix(iter, b, node_iter, k,
- k->u64s, k->u64s);
- if (bkey_whiteout(&insert->k)) {
- reserve_whiteout(b, k);
- return true;
+ insert->k.needs_whiteout = k->needs_whiteout;
+ k->needs_whiteout = false;
+
+ if (k >= btree_bset_last(b)->start) {
+ clobber_u64s = k->u64s;
+ goto overwrite;
} else {
- k->needs_whiteout = false;
+ bch2_btree_iter_fix_key_modified(iter, b, k);
}
- } else {
- /*
- * Deleting, but the key to delete wasn't found - nothing to do:
- */
- if (bkey_whiteout(&insert->k))
- return false;
-
- insert->k.needs_whiteout = false;
}
k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
- clobber_u64s = 0;
overwrite:
bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
- bch2_btree_node_iter_fix(iter, b, node_iter, k,
- clobber_u64s, k->u64s);
+ new_u64s = k->u64s;
+fix_iter:
+ if (clobber_u64s != new_u64s)
+ bch2_btree_node_iter_fix(iter, b, node_iter, k,
+ clobber_u64s, new_u64s);
return true;
}
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)
return __btree_node_flush(j, pin, 1, seq);
}
-static inline void __btree_journal_key(struct btree_trans *trans,
- enum btree_id btree_id,
- struct bkey_i *insert)
+inline void bch2_btree_add_journal_pin(struct bch_fs *c,
+ struct btree *b, u64 seq)
{
- 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;
-}
-
-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[0].b;
struct btree_write *w = btree_current_write(b);
- EBUG_ON(iter->level || b->level);
- 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);
- }
-
- if (unlikely(!journal_pin_active(&w->journal))) {
- u64 seq = likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
- ? trans->journal_res.seq
- : j->replay_journal_seq;
-
- bch2_journal_pin_add(j, seq, &w->journal,
- btree_node_write_idx(b) == 0
- ? btree_node_flush0
- : btree_node_flush1);
- }
-
- if (unlikely(!btree_node_dirty(b)))
- set_btree_node_dirty(b);
-}
-
-static void bch2_insert_fixup_key(struct btree_trans *trans,
- struct btree_insert_entry *insert)
-{
- struct btree_iter *iter = insert->iter;
- struct btree_iter_level *l = &iter->l[0];
-
- EBUG_ON(iter->level);
- EBUG_ON(insert->k->k.u64s >
- bch_btree_keys_u64s_remaining(trans->c, l->b));
-
- if (likely(bch2_btree_bset_insert_key(iter, l->b, &l->iter,
- insert->k)))
- bch2_btree_journal_key(trans, iter, insert->k);
+ bch2_journal_pin_add(&c->journal, seq, &w->journal,
+ btree_node_write_idx(b) == 0
+ ? btree_node_flush0
+ : btree_node_flush1);
}
/**
* btree_insert_key - insert a key one key into a leaf node
*/
-static void btree_insert_key_leaf(struct btree_trans *trans,
- struct btree_insert_entry *insert)
+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_iter *iter = insert->iter;
- struct btree *b = iter->l[0].b;
+ 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;
- if (!btree_node_is_extents(b))
- bch2_insert_fixup_key(trans, insert);
- else
- bch2_insert_fixup_extent(trans, insert);
+ 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;
+
+ 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(b);
live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
u64s_added = (int) bset_u64s(t) - old_u64s;
bch2_maybe_compact_whiteouts(c, b))
bch2_btree_iter_reinit_node(iter, b);
- trace_btree_insert_key(c, b, insert->k);
+ 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 btree_insert_entry *i)
+ struct btree_iter *iter,
+ struct bkey_i *insert)
{
struct bch_fs *c = trans->c;
- BUG_ON(i->iter->level);
- BUG_ON(bkey_cmp(bkey_start_pos(&i->k->k), i->iter->pos));
- EBUG_ON((i->iter->flags & BTREE_ITER_IS_EXTENTS) &&
- bkey_cmp(i->k->k.p, i->iter->l[0].b->key.k.p) > 0);
- EBUG_ON((i->iter->flags & BTREE_ITER_IS_EXTENTS) &&
- !(trans->flags & BTREE_INSERT_ATOMIC));
-
+ BUG_ON(bkey_cmp(insert->k.p, iter->pos));
BUG_ON(debug_check_bkeys(c) &&
- !bkey_deleted(&i->k->k) &&
- bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->iter->btree_id));
+ 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_insert_entry *insert,
+ struct btree_iter *iter,
+ struct bkey_i *insert,
unsigned *u64s)
{
struct bch_fs *c = trans->c;
- struct btree *b = insert->iter->l[0].b;
+ struct btree *b = iter_l(iter)->b;
static enum btree_insert_ret ret;
if (unlikely(btree_node_fake(b)))
return BTREE_INSERT_BTREE_NODE_FULL;
- ret = !btree_node_is_extents(b)
+ /*
+ * 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;
+
+ ret = !(iter->flags & BTREE_ITER_IS_EXTENTS)
? BTREE_INSERT_OK
- : bch2_extent_can_insert(trans, insert, u64s);
+ : bch2_extent_can_insert(trans, iter, insert);
if (ret)
return ret;
return BTREE_INSERT_OK;
}
+static enum btree_insert_ret
+btree_key_can_insert_cached(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_i *insert,
+ unsigned *u64s)
+{
+ struct bkey_cached *ck = (void *) iter->l[0].b;
+ unsigned new_u64s;
+ struct bkey_i *new_k;
+
+ BUG_ON(iter->level);
+
+ 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;
+}
+
static inline void do_btree_insert_one(struct btree_trans *trans,
- struct btree_insert_entry *insert)
+ struct btree_iter *iter,
+ struct bkey_i *insert)
{
- btree_insert_key_leaf(trans, 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 update_has_trans_triggers(struct btree_insert_entry *i)
+static inline bool iter_has_trans_triggers(struct btree_iter *iter)
{
- return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->iter->btree_id);
+ return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id);
}
-static inline bool update_has_nontrans_triggers(struct btree_insert_entry *i)
+static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
{
return (BTREE_NODE_TYPE_HAS_TRIGGERS &
~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) &
- (1U << i->iter->btree_id);
+ (1U << iter->btree_id);
}
static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
{
struct bch_fs *c = trans->c;
struct btree_insert_entry *i;
- unsigned mark_flags = trans->flags & BTREE_INSERT_BUCKET_INVALIDATE
- ? BCH_BUCKET_MARK_BUCKET_INVALIDATE
- : 0;
- if (unlikely(trans->flags & BTREE_INSERT_NOMARK))
- return;
+ trans_for_each_update(trans, i) {
+ /*
+ * XXX: synchronization of cached update triggers with gc
+ */
+ BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
- trans_for_each_update(trans, i)
if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
- bch2_mark_update(trans, i, NULL,
- mark_flags|BCH_BUCKET_MARK_GC);
+ bch2_mark_update(trans, i->iter, i->k, NULL,
+ i->trigger_flags|BTREE_TRIGGER_GC);
+ }
}
static inline int
struct bch_fs *c = trans->c;
struct bch_fs_usage *fs_usage = NULL;
struct btree_insert_entry *i;
- unsigned mark_flags = trans->flags & BTREE_INSERT_BUCKET_INVALIDATE
- ? BCH_BUCKET_MARK_BUCKET_INVALIDATE
- : 0;
- unsigned iter, u64s = 0;
+ unsigned u64s = 0;
bool marking = false;
int ret;
prefetch(&trans->c->journal.flags);
- trans_for_each_update_sorted(trans, i, iter) {
+ trans_for_each_update2(trans, i) {
/* Multiple inserts might go to same leaf: */
- if (!same_leaf_as_prev(trans, iter))
+ if (!same_leaf_as_prev(trans, i))
u64s = 0;
u64s += i->k->k.u64s;
- ret = btree_key_can_insert(trans, i, &u64s);
+ ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
+ ? btree_key_can_insert(trans, i->iter, i->k, &u64s)
+ : btree_key_can_insert_cached(trans, i->iter, i->k, &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))
- trans_for_each_update(trans, i)
+ trans_for_each_update2(trans, i)
i->k->k.version.lo = trans->journal_res.seq;
else if (inject_invalid_keys(c))
- trans_for_each_update(trans, i)
+ trans_for_each_update2(trans, i)
i->k->k.version = MAX_VERSION;
}
}
trans_for_each_update(trans, i)
- if (likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
- update_has_nontrans_triggers(i))
- bch2_mark_update(trans, i, fs_usage, mark_flags);
+ if (iter_has_nontrans_triggers(i->iter))
+ bch2_mark_update(trans, i->iter, i->k,
+ fs_usage, i->trigger_flags);
if (marking)
bch2_trans_fs_usage_apply(trans, fs_usage);
if (unlikely(c->gc_pos.phase))
bch2_trans_mark_gc(trans);
- trans_for_each_update(trans, i)
- do_btree_insert_one(trans, i);
+ trans_for_each_update2(trans, i)
+ do_btree_insert_one(trans, i->iter, i->k);
err:
if (marking) {
bch2_fs_usage_scratch_put(c, fs_usage);
{
struct btree_insert_entry *i;
struct btree_iter *iter;
- unsigned idx, u64s, journal_preres_u64s = 0;
int ret;
- /*
- * note: running triggers will append more updates to the list of
- * updates as we're walking it:
- */
- trans_for_each_update(trans, i) {
- /* we know trans->nounlock won't be set here: */
- if (unlikely(!(i->iter->locks_want < 1
- ? __bch2_btree_iter_upgrade(i->iter, 1)
- : i->iter->uptodate <= BTREE_ITER_NEED_PEEK))) {
- trace_trans_restart_upgrade(trans->ip);
- return -EINTR;
- }
-
- if (likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
- update_has_trans_triggers(i)) {
- ret = bch2_trans_mark_update(trans, i->iter, i->k);
- if (unlikely(ret)) {
- if (ret == -EINTR)
- trace_trans_restart_mark(trans->ip);
- return ret;
- }
- }
-
- u64s = jset_u64s(i->k->k.u64s);
- if (0)
- journal_preres_u64s += u64s;
- trans->journal_u64s += u64s;
- }
+ trans_for_each_update2(trans, i)
+ BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
ret = bch2_journal_preres_get(&trans->c->journal,
- &trans->journal_preres, journal_preres_u64s,
- JOURNAL_RES_GET_NONBLOCK);
+ &trans->journal_preres, trans->journal_preres_u64s,
+ 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,
- journal_preres_u64s);
+ trans->journal_preres_u64s);
if (unlikely(ret))
return ret;
* 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));
}
if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
- trans_for_each_update(trans, i)
- btree_insert_entry_checks(trans, i);
+ trans_for_each_update2(trans, i)
+ btree_insert_entry_checks(trans, i->iter, i->k);
bch2_btree_trans_verify_locks(trans);
- /*
- * No more updates can be added - sort updates so we can take write
- * locks in the correct order:
- */
- btree_trans_sort_updates(trans);
-
- trans_for_each_update_sorted(trans, i, idx)
- if (!same_leaf_as_prev(trans, idx))
+ trans_for_each_update2(trans, i)
+ if (!same_leaf_as_prev(trans, i))
bch2_btree_node_lock_for_insert(trans->c,
- i->iter->l[0].b, i->iter);
+ iter_l(i->iter)->b, i->iter);
ret = bch2_trans_commit_write_locked(trans, stopped_at);
- trans_for_each_update_sorted(trans, i, idx)
- if (!same_leaf_as_prev(trans, idx))
- bch2_btree_node_unlock_write_inlined(i->iter->l[0].b,
+ trans_for_each_update2(trans, i)
+ if (!same_leaf_as_prev(trans, i))
+ 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:
if (trans->flags & BTREE_INSERT_NOUNLOCK)
trans->nounlock = true;
- trans_for_each_update_sorted(trans, i, idx)
- if (!same_leaf_as_prev(trans, idx))
+ trans_for_each_update2(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_update(trans, i)
- bch2_btree_iter_downgrade(i->iter);
+ bch2_trans_downgrade(trans);
return 0;
}
/*
* if the split succeeded without dropping locks the insert will
- * still be atomic (in the BTREE_INSERT_ATOMIC sense, what the
- * caller peeked() and is overwriting won't have changed)
+ * still be atomic (what the caller peeked() and is overwriting
+ * won't have changed)
*/
#if 0
/*
return ret2;
}
- /*
- * BTREE_ITER_ATOMIC means we have to return -EINTR if we
- * dropped locks:
- */
- if (!(flags & BTREE_INSERT_ATOMIC))
- return 0;
-
trace_trans_restart_atomic(trans->ip);
}
return 0;
}
+static void bch2_trans_update2(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_i *insert)
+{
+ struct btree_insert_entry *i, n = (struct btree_insert_entry) {
+ .iter = iter, .k = insert
+ };
+
+ btree_insert_entry_checks(trans, n.iter, n.k);
+
+ BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
+
+ EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
+
+ iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+
+ trans_for_each_update2(trans, i) {
+ if (btree_iter_cmp(n.iter, i->iter) == 0) {
+ *i = n;
+ return;
+ }
+
+ if (btree_iter_cmp(n.iter, i->iter) <= 0)
+ break;
+ }
+
+ array_insert_item(trans->updates2, trans->nr_updates2,
+ i - trans->updates2, n);
+}
+
+static int extent_update_to_keys(struct btree_trans *trans,
+ struct btree_iter *orig_iter,
+ struct bkey_i *insert)
+{
+ struct btree_iter *iter;
+
+ if (bkey_deleted(&insert->k))
+ return 0;
+
+ iter = bch2_trans_copy_iter(trans, orig_iter);
+ if (IS_ERR(iter))
+ return PTR_ERR(iter);
+
+ iter->flags |= BTREE_ITER_INTENT;
+ __bch2_btree_iter_set_pos(iter, insert->k.p, false);
+ bch2_trans_update2(trans, iter, insert);
+ bch2_trans_iter_put(trans, iter);
+ return 0;
+}
+
+static int extent_handle_overwrites(struct btree_trans *trans,
+ enum btree_id btree_id,
+ struct bpos start, struct bpos end)
+{
+ struct btree_iter *iter = NULL, *update_iter;
+ struct bkey_i *update;
+ struct bkey_s_c k;
+ int ret = 0;
+
+ iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
+ ret = PTR_ERR_OR_ZERO(iter);
+ if (ret)
+ return ret;
+
+ k = bch2_btree_iter_peek_with_updates(iter);
+
+ while (k.k && !(ret = bkey_err(k))) {
+ if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
+ break;
+
+ if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
+ update_iter = bch2_trans_copy_iter(trans, iter);
+ if ((ret = PTR_ERR_OR_ZERO(update_iter)))
+ goto err;
+
+ update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ goto err;
+
+ bkey_reassemble(update, k);
+ bch2_cut_back(start, update);
+
+ __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
+ bch2_trans_update2(trans, update_iter, update);
+ bch2_trans_iter_put(trans, update_iter);
+ }
+
+ if (bkey_cmp(k.k->p, end) > 0) {
+ update_iter = bch2_trans_copy_iter(trans, iter);
+ if ((ret = PTR_ERR_OR_ZERO(update_iter)))
+ goto err;
+
+ update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ goto err;
+
+ bkey_reassemble(update, k);
+ bch2_cut_front(end, update);
+
+ __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
+ bch2_trans_update2(trans, update_iter, update);
+ bch2_trans_iter_put(trans, update_iter);
+ } else {
+ update_iter = bch2_trans_copy_iter(trans, iter);
+ if ((ret = PTR_ERR_OR_ZERO(update_iter)))
+ goto err;
+
+ update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
+ if ((ret = PTR_ERR_OR_ZERO(update)))
+ goto err;
+
+ update->k = *k.k;
+ set_bkey_val_u64s(&update->k, 0);
+ update->k.type = KEY_TYPE_deleted;
+ update->k.size = 0;
+
+ __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
+ bch2_trans_update2(trans, update_iter, update);
+ bch2_trans_iter_put(trans, update_iter);
+ }
+
+ k = bch2_btree_iter_next_with_updates(iter);
+ }
+err:
+ if (!IS_ERR_OR_NULL(iter))
+ bch2_trans_iter_put(trans, iter);
+ return ret;
+}
+
int __bch2_trans_commit(struct btree_trans *trans)
{
struct btree_insert_entry *i = NULL;
struct btree_iter *iter;
- unsigned orig_nr_updates = trans->nr_updates;
- unsigned orig_mem_top = trans->mem_top;
+ bool trans_trigger_run;
+ unsigned u64s;
int ret = 0;
+ BUG_ON(trans->need_reset);
+
if (!trans->nr_updates)
goto out_noupdates;
- /* for the sake of sanity: */
- EBUG_ON(trans->nr_updates > 1 && !(trans->flags & BTREE_INSERT_ATOMIC));
-
if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
lockdep_assert_held(&trans->c->gc_lock);
memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
+ trans->journal_u64s = trans->extra_journal_entry_u64s;
+ trans->journal_preres_u64s = 0;
+
if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
unlikely(!percpu_ref_tryget(&trans->c->writes))) {
ret = bch2_trans_commit_get_rw_cold(trans);
if (ret)
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:
+ */
+ do {
+ trans_trigger_run = false;
+
+ trans_for_each_update(trans, i) {
+ if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK &&
+ (ret = bch2_btree_iter_traverse(i->iter)))) {
+ trace_trans_restart_traverse(trans->ip);
+ goto out;
+ }
+
+ /*
+ * We're not using bch2_btree_iter_upgrade here because
+ * we know trans->nounlock can't be set:
+ */
+ if (unlikely(i->iter->locks_want < 1 &&
+ !__bch2_btree_iter_upgrade(i->iter, 1))) {
+ trace_trans_restart_upgrade(trans->ip);
+ ret = -EINTR;
+ goto out;
+ }
+
+ if (iter_has_trans_triggers(i->iter) &&
+ !i->trans_triggers_run) {
+ i->trans_triggers_run = true;
+ trans_trigger_run = true;
+
+ ret = bch2_trans_mark_update(trans, i->iter, i->k,
+ i->trigger_flags);
+ if (unlikely(ret)) {
+ if (ret == -EINTR)
+ trace_trans_restart_mark(trans->ip);
+ goto out;
+ }
+ }
+ }
+ } while (trans_trigger_run);
+
+ /* Turn extents updates into keys: */
+ trans_for_each_update(trans, i)
+ if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
+ struct bpos start = bkey_start_pos(&i->k->k);
+
+ while (i + 1 < trans->updates + trans->nr_updates &&
+ i[0].iter->btree_id == i[1].iter->btree_id &&
+ !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
+ i++;
+
+ ret = extent_handle_overwrites(trans, i->iter->btree_id,
+ start, i->k->k.p);
+ if (ret)
+ goto out;
+ }
+
+ trans_for_each_update(trans, i) {
+ if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
+ ret = extent_update_to_keys(trans, i->iter, i->k);
+ if (ret)
+ goto out;
+ } else {
+ bch2_trans_update2(trans, i->iter, i->k);
+ }
+ }
+
+ trans_for_each_update2(trans, i) {
+ BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
+ BUG_ON(i->iter->locks_want < 1);
+
+ u64s = jset_u64s(i->k->k.u64s);
+ 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;
+ }
retry:
memset(&trans->journal_res, 0, sizeof(trans->journal_res));
- trans->journal_u64s = 0;
ret = do_bch2_trans_commit(trans, &i);
- if (trans->fs_usage_deltas) {
- trans->fs_usage_deltas->used = 0;
- memset(&trans->fs_usage_deltas->memset_start, 0,
- (void *) &trans->fs_usage_deltas->memset_end -
- (void *) &trans->fs_usage_deltas->memset_start);
- }
-
/* make sure we didn't drop or screw up locks: */
bch2_btree_trans_verify_locks(trans);
if (ret)
goto err;
+
+ trans_for_each_iter(trans, iter)
+ if ((trans->iters_live & (1ULL << iter->idx)) &&
+ (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
+ if (trans->flags & BTREE_INSERT_NOUNLOCK)
+ bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
+ else
+ bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
+ }
out:
bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
percpu_ref_put(&trans->c->writes);
out_noupdates:
- EBUG_ON(!(trans->flags & BTREE_INSERT_ATOMIC) && ret == -EINTR);
-
- trans_for_each_iter_all(trans, iter)
- iter->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
-
- if (!ret) {
- bch2_trans_unlink_iters(trans);
- trans->iters_touched = 0;
- }
- trans->nr_updates = 0;
- trans->mem_top = 0;
+ bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
return ret;
err:
ret = bch2_trans_commit_error(trans, i, ret);
-
- /* can't loop if it was passed in and we changed it: */
- if (unlikely(trans->flags & BTREE_INSERT_NO_CLEAR_REPLICAS) && !ret)
- ret = -EINTR;
if (ret)
goto out;
- /* free updates and memory used by triggers, they'll be reexecuted: */
- trans->nr_updates = orig_nr_updates;
- trans->mem_top = orig_mem_top;
goto retry;
}
+int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
+ struct bkey_i *k, enum btree_trigger_flags flags)
+{
+ struct btree_insert_entry *i, n = (struct btree_insert_entry) {
+ .trigger_flags = flags, .iter = iter, .k = k
+ };
+
+ EBUG_ON(bkey_cmp(iter->pos,
+ (iter->flags & BTREE_ITER_IS_EXTENTS)
+ ? bkey_start_pos(&k->k)
+ : k->k.p));
+
+ iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+
+ if (btree_node_type_is_extents(iter->btree_id)) {
+ iter->pos_after_commit = k->k.p;
+ iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
+ }
+
+ /*
+ * 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)
+ break;
+
+ /*
+ * Now delete/trim any updates the new update overwrites:
+ */
+ if (i > trans->updates &&
+ i[-1].iter->btree_id == iter->btree_id &&
+ bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
+ bch2_cut_back(n.iter->pos, i[-1].k);
+
+ while (i < trans->updates + trans->nr_updates &&
+ iter->btree_id == i->iter->btree_id &&
+ bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
+ array_remove_item(trans->updates, trans->nr_updates,
+ i - trans->updates);
+
+ if (i < trans->updates + trans->nr_updates &&
+ iter->btree_id == i->iter->btree_id &&
+ bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
+ /*
+ * When we have an extent that overwrites the start of another
+ * update, trimming that extent will mean the iterator's
+ * position has to change since the iterator position has to
+ * match the extent's start pos - but we don't want to change
+ * the iterator pos if some other code is using it, so we may
+ * need to clone it:
+ */
+ if (trans->iters_live & (1ULL << i->iter->idx)) {
+ i->iter = bch2_trans_copy_iter(trans, i->iter);
+ if (IS_ERR(i->iter)) {
+ trans->need_reset = true;
+ return PTR_ERR(i->iter);
+ }
+
+ i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+ bch2_trans_iter_put(trans, i->iter);
+ }
+
+ bch2_cut_front(n.k->k.p, i->k);
+ bch2_btree_iter_set_pos(i->iter, n.k->k.p);
+ }
+
+ EBUG_ON(trans->nr_updates >= trans->nr_iters);
+
+ array_insert_item(trans->updates, trans->nr_updates,
+ i - trans->updates, n);
+ return 0;
+}
+
+int __bch2_btree_insert(struct btree_trans *trans,
+ enum btree_id id, struct bkey_i *k)
+{
+ struct btree_iter *iter;
+ int ret;
+
+ iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
+ BTREE_ITER_INTENT);
+ if (IS_ERR(iter))
+ return PTR_ERR(iter);
+
+ ret = bch2_btree_iter_traverse(iter) ?:
+ bch2_trans_update(trans, iter, k, 0);
+ bch2_trans_iter_put(trans, iter);
+ return ret;
+}
+
/**
* bch2_btree_insert - insert keys into the extent btree
* @c: pointer to struct bch_fs
* @hook: insert callback
*/
int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
- struct bkey_i *k,
- struct disk_reservation *disk_res,
- u64 *journal_seq, int flags)
+ struct bkey_i *k,
+ struct disk_reservation *disk_res,
+ u64 *journal_seq, int flags)
{
- struct btree_trans trans;
- struct btree_iter *iter;
- int ret;
-
- bch2_trans_init(&trans, c, 0, 0);
-retry:
- bch2_trans_begin(&trans);
-
- iter = bch2_trans_get_iter(&trans, id, bkey_start_pos(&k->k),
- BTREE_ITER_INTENT);
-
- bch2_trans_update(&trans, iter, k);
-
- ret = bch2_trans_commit(&trans, disk_res, journal_seq, flags);
- if (ret == -EINTR)
- goto retry;
- bch2_trans_exit(&trans);
-
- return ret;
+ return bch2_trans_do(c, disk_res, journal_seq, flags,
+ __bch2_btree_insert(&trans, id, k));
}
int bch2_btree_delete_at_range(struct btree_trans *trans,
bkey_cmp(iter->pos, end) < 0) {
struct bkey_i delete;
- bch2_trans_unlink_iters(trans);
- trans->iters_touched &= trans->iters_live;
+ bch2_trans_begin(trans);
bkey_init(&delete.k);
*/
delete.k.p = iter->pos;
- if (iter->flags & BTREE_ITER_IS_EXTENTS) {
+ if (btree_node_type_is_extents(iter->btree_id)) {
unsigned max_sectors =
KEY_SIZE_MAX & (~0 << trans->c->block_bits);
break;
}
- bch2_trans_update(trans, iter, &delete);
+ bch2_trans_update(trans, iter, &delete, 0);
ret = bch2_trans_commit(trans, NULL, journal_seq,
- BTREE_INSERT_ATOMIC|
BTREE_INSERT_NOFAIL);
if (ret)
break;
bkey_init(&k.k);
k.k.p = iter->pos;
- bch2_trans_update(trans, iter, &k);
+ bch2_trans_update(trans, iter, &k, 0);
return bch2_trans_commit(trans, NULL, NULL,
BTREE_INSERT_NOFAIL|
BTREE_INSERT_USE_RESERVE|flags);