#include "bcachefs.h"
#include "btree_update.h"
#include "btree_update_interior.h"
+#include "btree_gc.h"
#include "btree_io.h"
#include "btree_iter.h"
#include "btree_locking.h"
#include <linux/sort.h>
#include <trace/events/bcachefs.h>
-static bool btree_trans_relock(struct btree_insert *);
-static void btree_trans_unlock(struct btree_insert *);
+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_node_just_written(b) &&
+ bch2_btree_post_write_cleanup(c, b))
+ bch2_btree_iter_reinit_node(iter, b);
+
+ /*
+ * If the last bset has been written, or if it's gotten too big - start
+ * a new bset to insert into:
+ */
+ if (want_new_bset(c, b))
+ bch2_btree_init_next(c, b, iter);
+}
+
+static void btree_trans_lock_write(struct bch_fs *c, struct btree_trans *trans)
+{
+ struct btree_insert_entry *i;
+
+ trans_for_each_update_leaf(trans, i)
+ bch2_btree_node_lock_for_insert(c, i->iter->l[0].b, i->iter);
+}
+
+static void btree_trans_unlock_write(struct btree_trans *trans)
+{
+ struct btree_insert_entry *i;
+
+ trans_for_each_update_leaf(trans, i)
+ bch2_btree_node_unlock_write(i->iter->l[0].b, i->iter);
+}
+
+static inline int btree_trans_cmp(struct btree_insert_entry l,
+ struct btree_insert_entry r)
+{
+ return cmp_int(l.deferred, r.deferred) ?:
+ btree_iter_cmp(l.iter, r.iter);
+}
/* Inserting into a given leaf node (last stage of insert): */
return __btree_node_flush(j, pin, 1, seq);
}
-static inline void __btree_journal_key(struct btree_insert *trans,
+static inline void __btree_journal_key(struct btree_trans *trans,
enum btree_id btree_id,
struct bkey_i *insert)
{
*trans->journal_seq = seq;
}
-void bch2_btree_journal_key(struct btree_insert *trans,
+void bch2_btree_journal_key(struct btree_trans *trans,
struct btree_iter *iter,
struct bkey_i *insert)
{
set_btree_node_dirty(b);
}
-static enum btree_insert_ret
-bch2_insert_fixup_key(struct btree_insert *trans,
- struct btree_insert_entry *insert)
+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];
if (bch2_btree_bset_insert_key(iter, l->b, &l->iter,
insert->k))
bch2_btree_journal_key(trans, iter, insert->k);
-
- return BTREE_INSERT_OK;
}
/**
* btree_insert_key - insert a key one key into a leaf node
*/
-static enum btree_insert_ret
-btree_insert_key_leaf(struct btree_insert *trans,
- struct btree_insert_entry *insert)
+static void btree_insert_key_leaf(struct btree_trans *trans,
+ struct btree_insert_entry *insert)
{
struct bch_fs *c = trans->c;
struct btree_iter *iter = insert->iter;
struct btree *b = iter->l[0].b;
- enum btree_insert_ret ret;
int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
int old_live_u64s = b->nr.live_u64s;
int live_u64s_added, u64s_added;
- bch2_mark_update(trans, insert);
-
- ret = !btree_node_is_extents(b)
- ? bch2_insert_fixup_key(trans, insert)
- : bch2_insert_fixup_extent(trans, insert);
+ if (!btree_node_is_extents(b))
+ bch2_insert_fixup_key(trans, insert);
+ else
+ bch2_insert_fixup_extent(trans, insert);
live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
bch2_btree_iter_reinit_node(iter, b);
trace_btree_insert_key(c, b, insert->k);
- return ret;
}
/* Deferred btree updates: */
kfree(k);
}
-static enum btree_insert_ret
-btree_insert_key_deferred(struct btree_insert *trans,
- struct btree_insert_entry *insert)
+static void btree_insert_key_deferred(struct btree_trans *trans,
+ struct btree_insert_entry *insert)
{
struct bch_fs *c = trans->c;
struct journal *j = &c->journal;
bch2_journal_pin_update(j, trans->journal_res.seq, &d->journal,
deferred_update_flush);
spin_unlock(&d->lock);
-
- return BTREE_INSERT_OK;
}
void bch2_deferred_update_free(struct bch_fs *c,
return d;
}
-/* struct btree_insert operations: */
+/* Normal update interface: */
-/*
- * We sort transaction entries so that if multiple iterators point to the same
- * leaf node they'll be adjacent:
- */
-static bool same_leaf_as_prev(struct btree_insert *trans,
- struct btree_insert_entry *i)
+static inline void btree_insert_entry_checks(struct btree_trans *trans,
+ struct btree_insert_entry *i)
{
- return i != trans->entries &&
- !i->deferred &&
- i[0].iter->l[0].b == i[-1].iter->l[0].b;
-}
-
-#define __trans_next_entry(_trans, _i, _filter) \
-({ \
- while ((_i) < (_trans)->entries + (_trans->nr) && !(_filter)) \
- (_i)++; \
- \
- (_i) < (_trans)->entries + (_trans->nr); \
-})
-
-#define __trans_for_each_entry(_trans, _i, _filter) \
- for ((_i) = (_trans)->entries; \
- __trans_next_entry(_trans, _i, _filter); \
- (_i)++)
+ struct bch_fs *c = trans->c;
+ enum btree_id btree_id = !i->deferred
+ ? i->iter->btree_id
+ : i->d->btree_id;
-#define trans_for_each_entry(trans, i) \
- __trans_for_each_entry(trans, i, true)
+ if (!i->deferred) {
+ 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) &&
+ !bch2_extent_is_atomic(i->k, i->iter));
-#define trans_for_each_iter(trans, i) \
- __trans_for_each_entry(trans, i, !(i)->deferred)
+ EBUG_ON((i->iter->flags & BTREE_ITER_IS_EXTENTS) &&
+ !(trans->flags & BTREE_INSERT_ATOMIC));
+ }
-#define trans_for_each_leaf(trans, i) \
- __trans_for_each_entry(trans, i, !(i)->deferred && \
- !same_leaf_as_prev(trans, i))
+ BUG_ON(debug_check_bkeys(c) &&
+ !bkey_deleted(&i->k->k) &&
+ bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), btree_id));
+}
-inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
- struct btree_iter *iter)
+static int bch2_trans_journal_preres_get(struct btree_trans *trans)
{
- bch2_btree_node_lock_write(b, iter);
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i;
+ unsigned u64s = 0;
+ int ret;
- if (btree_node_just_written(b) &&
- bch2_btree_post_write_cleanup(c, b))
- bch2_btree_iter_reinit_node(iter, b);
+ trans_for_each_update(trans, i)
+ if (i->deferred)
+ u64s += jset_u64s(i->k->k.u64s);
- /*
- * If the last bset has been written, or if it's gotten too big - start
- * a new bset to insert into:
- */
- if (want_new_bset(c, b))
- bch2_btree_init_next(c, b, iter);
-}
+ if (!u64s)
+ return 0;
-static void multi_lock_write(struct bch_fs *c, struct btree_insert *trans)
-{
- struct btree_insert_entry *i;
+ ret = bch2_journal_preres_get(&c->journal,
+ &trans->journal_preres, u64s,
+ JOURNAL_RES_GET_NONBLOCK);
+ if (ret != -EAGAIN)
+ return ret;
- trans_for_each_leaf(trans, i)
- bch2_btree_node_lock_for_insert(c, i->iter->l[0].b, i->iter);
-}
+ bch2_trans_unlock(trans);
-static void multi_unlock_write(struct btree_insert *trans)
-{
- struct btree_insert_entry *i;
+ ret = bch2_journal_preres_get(&c->journal,
+ &trans->journal_preres, u64s, 0);
+ if (ret)
+ return ret;
- trans_for_each_leaf(trans, i)
- bch2_btree_node_unlock_write(i->iter->l[0].b, i->iter);
-}
+ if (!bch2_trans_relock(trans)) {
+ trace_trans_restart_journal_preres_get(trans->ip);
+ return -EINTR;
+ }
-static inline int btree_trans_cmp(struct btree_insert_entry l,
- struct btree_insert_entry r)
-{
- return (l.deferred > r.deferred) - (l.deferred < r.deferred) ?:
- btree_iter_cmp(l.iter, r.iter);
+ return 0;
}
-static bool btree_trans_relock(struct btree_insert *trans)
+static int bch2_trans_journal_res_get(struct btree_trans *trans,
+ unsigned flags)
{
- struct btree_insert_entry *i;
+ struct bch_fs *c = trans->c;
+ int ret;
- trans_for_each_iter(trans, i)
- return bch2_btree_iter_relock(i->iter);
- return true;
-}
+ if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
+ flags |= JOURNAL_RES_GET_RESERVED;
-static void btree_trans_unlock(struct btree_insert *trans)
-{
- struct btree_insert_entry *i;
+ ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
+ trans->journal_u64s, flags);
- trans_for_each_iter(trans, i) {
- bch2_btree_iter_unlock(i->iter);
- break;
- }
+ return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
}
-/* Normal update interface: */
-
static enum btree_insert_ret
-btree_key_can_insert(struct btree_insert *trans,
+btree_key_can_insert(struct btree_trans *trans,
struct btree_insert_entry *insert,
unsigned *u64s)
{
if (unlikely(btree_node_fake(b)))
return BTREE_INSERT_BTREE_NODE_FULL;
- if (!bch2_bkey_replicas_marked(c,
- bkey_i_to_s_c(insert->k),
- true))
- return BTREE_INSERT_NEED_MARK_REPLICAS;
-
ret = !btree_node_is_extents(b)
? BTREE_INSERT_OK
: bch2_extent_can_insert(trans, insert, u64s);
return BTREE_INSERT_OK;
}
-static inline enum btree_insert_ret
-do_btree_insert_one(struct btree_insert *trans,
- struct btree_insert_entry *insert)
+static int btree_trans_check_can_insert(struct btree_trans *trans,
+ struct btree_insert_entry **stopped_at)
+{
+ struct btree_insert_entry *i;
+ unsigned u64s = 0;
+ int ret;
+
+ trans_for_each_update_iter(trans, i) {
+ /* Multiple inserts might go to same leaf: */
+ if (!same_leaf_as_prev(trans, i))
+ u64s = 0;
+
+ u64s += i->k->k.u64s;
+ ret = btree_key_can_insert(trans, i, &u64s);
+ if (ret) {
+ *stopped_at = i;
+ return ret;
+ }
+ }
+
+ return 0;
+}
+
+static inline void do_btree_insert_one(struct btree_trans *trans,
+ struct btree_insert_entry *insert)
+{
+ if (likely(!insert->deferred))
+ btree_insert_key_leaf(trans, insert);
+ else
+ btree_insert_key_deferred(trans, insert);
+}
+
+static inline bool update_triggers_transactional(struct btree_trans *trans,
+ struct btree_insert_entry *i)
{
- return likely(!insert->deferred)
- ? btree_insert_key_leaf(trans, insert)
- : btree_insert_key_deferred(trans, insert);
+ return likely(!(trans->flags & BTREE_INSERT_MARK_INMEM)) &&
+ (i->iter->btree_id == BTREE_ID_EXTENTS ||
+ i->iter->btree_id == BTREE_ID_INODES);
+}
+
+static inline bool update_has_triggers(struct btree_trans *trans,
+ struct btree_insert_entry *i)
+{
+ return likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
+ !i->deferred &&
+ btree_node_type_needs_gc(i->iter->btree_id);
}
/*
* Get journal reservation, take write locks, and attempt to do btree update(s):
*/
-static inline int do_btree_insert_at(struct btree_insert *trans,
+static inline int do_btree_insert_at(struct btree_trans *trans,
struct btree_insert_entry **stopped_at)
{
struct bch_fs *c = trans->c;
+ struct bch_fs_usage *fs_usage = NULL;
struct btree_insert_entry *i;
- struct btree_iter *linked;
- unsigned u64s;
+ unsigned mark_flags = trans->flags & BTREE_INSERT_BUCKET_INVALIDATE
+ ? BCH_BUCKET_MARK_BUCKET_INVALIDATE
+ : 0;
int ret;
-retry:
- trans_for_each_iter(trans, i)
+
+ trans_for_each_update_iter(trans, i)
BUG_ON(i->iter->uptodate >= BTREE_ITER_NEED_RELOCK);
- memset(&trans->journal_res, 0, sizeof(trans->journal_res));
+ trans_for_each_update_iter(trans, i)
+ if (update_has_triggers(trans, i) &&
+ update_triggers_transactional(trans, i)) {
+ ret = bch2_trans_mark_update(trans, i);
+ if (ret == -EINTR)
+ trace_trans_restart_mark(trans->ip);
+ if (ret)
+ goto out_clear_replicas;
+ }
- multi_lock_write(c, trans);
+ btree_trans_lock_write(c, trans);
if (race_fault()) {
ret = -EINTR;
- trans_restart(" (race)");
+ trace_trans_restart_fault_inject(trans->ip);
goto out;
}
* held, otherwise another thread could write the node changing the
* amount of space available:
*/
- u64s = 0;
- trans_for_each_iter(trans, i) {
- /* Multiple inserts might go to same leaf: */
- if (!same_leaf_as_prev(trans, i))
- u64s = 0;
+ ret = btree_trans_check_can_insert(trans, stopped_at);
+ if (ret)
+ goto out;
- u64s += i->k->k.u64s;
- ret = btree_key_can_insert(trans, i, &u64s);
- if (ret) {
- *stopped_at = i;
+ trans_for_each_update_iter(trans, i) {
+ if (i->deferred ||
+ !btree_node_type_needs_gc(i->iter->btree_id))
+ continue;
+
+ if (!fs_usage) {
+ percpu_down_read(&c->mark_lock);
+ fs_usage = bch2_fs_usage_scratch_get(c);
+ }
+
+ if (!bch2_bkey_replicas_marked_locked(c,
+ bkey_i_to_s_c(i->k), true)) {
+ ret = BTREE_INSERT_NEED_MARK_REPLICAS;
goto out;
}
}
+ /*
+ * Don't get journal reservation until after we know insert will
+ * succeed:
+ */
if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
- unsigned flags = (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
- ? JOURNAL_RES_GET_RESERVED : 0;
-
- u64s = 0;
- trans_for_each_entry(trans, i)
- u64s += jset_u64s(i->k->k.u64s);
-
- ret = bch2_journal_res_get(&c->journal,
- &trans->journal_res, u64s,
- flags|JOURNAL_RES_GET_NONBLOCK);
- if (likely(!ret))
- goto got_journal_res;
- if (ret != -EAGAIN)
- goto out;
+ trans->journal_u64s = 0;
- multi_unlock_write(trans);
- btree_trans_unlock(trans);
+ trans_for_each_update(trans, i)
+ trans->journal_u64s += jset_u64s(i->k->k.u64s);
- ret = bch2_journal_res_get(&c->journal,
- &trans->journal_res, u64s,
- flags|JOURNAL_RES_GET_CHECK);
+ ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_NONBLOCK);
if (ret)
- return ret;
-
- if (!btree_trans_relock(trans)) {
- trans_restart(" (iter relock after journal res get blocked)");
- return -EINTR;
- }
-
- goto retry;
+ goto out;
}
-got_journal_res:
+
if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
if (journal_seq_verify(c))
- trans_for_each_entry(trans, i)
+ trans_for_each_update(trans, i)
i->k->k.version.lo = trans->journal_res.seq;
else if (inject_invalid_keys(c))
- trans_for_each_entry(trans, i)
+ trans_for_each_update(trans, i)
i->k->k.version = MAX_VERSION;
}
- if (trans->flags & BTREE_INSERT_NOUNLOCK) {
- /*
- * linked iterators that weren't being updated may or may not
- * have been traversed/locked, depending on what the caller was
- * doing:
- */
- trans_for_each_iter(trans, i) {
- for_each_btree_iter(i->iter, linked)
- if (linked->uptodate < BTREE_ITER_NEED_RELOCK)
- linked->flags |= BTREE_ITER_NOUNLOCK;
- break;
- }
- }
- trans->did_work = true;
+ trans_for_each_update_iter(trans, i)
+ if (update_has_triggers(trans, i) &&
+ !update_triggers_transactional(trans, i))
+ bch2_mark_update(trans, i, fs_usage, mark_flags);
- trans_for_each_entry(trans, i) {
- switch (do_btree_insert_one(trans, i)) {
- case BTREE_INSERT_OK:
- break;
- case BTREE_INSERT_NEED_TRAVERSE:
- BUG_ON((trans->flags &
- (BTREE_INSERT_ATOMIC|BTREE_INSERT_NOUNLOCK)));
- ret = -EINTR;
- goto out;
- default:
- BUG();
- }
- }
+ if (fs_usage && trans->fs_usage_deltas)
+ bch2_replicas_delta_list_apply(c, fs_usage,
+ trans->fs_usage_deltas);
+
+ if (fs_usage)
+ bch2_trans_fs_usage_apply(trans, fs_usage);
+
+ if (likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
+ unlikely(c->gc_pos.phase))
+ trans_for_each_update_iter(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);
+
+ trans_for_each_update(trans, i)
+ do_btree_insert_one(trans, i);
out:
BUG_ON(ret &&
(trans->flags & BTREE_INSERT_JOURNAL_RESERVED) &&
trans->journal_res.ref);
- multi_unlock_write(trans);
- bch2_journal_res_put(&c->journal, &trans->journal_res);
+ btree_trans_unlock_write(trans);
- return ret;
-}
-
-static inline void btree_insert_entry_checks(struct bch_fs *c,
- struct btree_insert_entry *i)
-{
- enum btree_id btree_id = !i->deferred
- ? i->iter->btree_id
- : i->d->btree_id;
-
- if (!i->deferred) {
- BUG_ON(i->iter->level);
- BUG_ON(bkey_cmp(bkey_start_pos(&i->k->k), i->iter->pos));
+ if (fs_usage) {
+ bch2_fs_usage_scratch_put(c, fs_usage);
+ percpu_up_read(&c->mark_lock);
+ }
- bch2_btree_iter_verify_locks(i->iter);
+ bch2_journal_res_put(&c->journal, &trans->journal_res);
+out_clear_replicas:
+ if (trans->fs_usage_deltas) {
+ memset(&trans->fs_usage_deltas->fs_usage, 0,
+ sizeof(trans->fs_usage_deltas->fs_usage));
+ trans->fs_usage_deltas->used = 0;
}
- BUG_ON(debug_check_bkeys(c) &&
- !bkey_deleted(&i->k->k) &&
- bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), btree_id));
+ return ret;
}
-/**
- * __bch_btree_insert_at - insert keys at given iterator positions
- *
- * This is main entry point for btree updates.
- *
- * Return values:
- * -EINTR: locking changed, this function should be called again. Only returned
- * if passed BTREE_INSERT_ATOMIC.
- * -EROFS: filesystem read only
- * -EIO: journal or btree node IO error
- */
-int __bch2_btree_insert_at(struct btree_insert *trans)
+static noinline
+int bch2_trans_commit_error(struct btree_trans *trans,
+ struct btree_insert_entry *i,
+ int ret)
{
struct bch_fs *c = trans->c;
- struct btree_insert_entry *i;
- struct btree_iter *linked;
- unsigned flags, u64s = 0;
- int ret;
-
- BUG_ON(!trans->nr);
+ unsigned flags = trans->flags;
+ struct btree_insert_entry *src, *dst;
- /* for the sake of sanity: */
- BUG_ON(trans->nr > 1 && !(trans->flags & BTREE_INSERT_ATOMIC));
-
- if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
- lockdep_assert_held(&c->gc_lock);
-
- memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
-
- bubble_sort(trans->entries, trans->nr, btree_trans_cmp);
-
- trans_for_each_entry(trans, i)
- btree_insert_entry_checks(c, i);
+ src = dst = trans->updates;
- trans_for_each_entry(trans, i)
- if (i->deferred)
- u64s += jset_u64s(i->k->k.u64s);
-
- if (u64s) {
- ret = bch2_journal_preres_get(&c->journal,
- &trans->journal_preres, u64s,
- JOURNAL_RES_GET_NONBLOCK);
- if (!ret)
- goto got_journal_preres;
- if (ret != -EAGAIN)
- return ret;
-
- btree_trans_unlock(trans);
- ret = bch2_journal_preres_get(&c->journal,
- &trans->journal_preres, u64s, 0);
- if (ret)
- return ret;
-
- if (!btree_trans_relock(trans)) {
- trans_restart(" (iter relock after journal preres get blocked)");
- bch2_journal_preres_put(&c->journal, &trans->journal_preres);
- return -EINTR;
+ while (src < trans->updates + trans->nr_updates) {
+ if (!src->triggered) {
+ *dst = *src;
+ dst++;
}
- }
-got_journal_preres:
- if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
- !percpu_ref_tryget(&c->writes)))
- return -EROFS;
-retry:
- trans_for_each_iter(trans, i) {
- unsigned old_locks_want = i->iter->locks_want;
- unsigned old_uptodate = i->iter->uptodate;
-
- if (!bch2_btree_iter_upgrade(i->iter, 1, true)) {
- trans_restart(" (failed upgrade, locks_want %u uptodate %u)",
- old_locks_want, old_uptodate);
- ret = -EINTR;
- goto err;
- }
-
- if (i->iter->flags & BTREE_ITER_ERROR) {
- ret = -EIO;
- goto err;
- }
- }
-
- ret = do_btree_insert_at(trans, &i);
- if (unlikely(ret))
- goto err;
-
- trans_for_each_leaf(trans, i)
- bch2_foreground_maybe_merge(c, i->iter, 0, trans->flags);
-
- trans_for_each_iter(trans, i)
- bch2_btree_iter_downgrade(i->iter);
-out:
- bch2_journal_preres_put(&c->journal, &trans->journal_preres);
-
- if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
- percpu_ref_put(&c->writes);
-
- /* make sure we didn't drop or screw up locks: */
- trans_for_each_iter(trans, i) {
- bch2_btree_iter_verify_locks(i->iter);
- break;
+ src++;
}
- trans_for_each_iter(trans, i) {
- for_each_btree_iter(i->iter, linked)
- linked->flags &= ~BTREE_ITER_NOUNLOCK;
- break;
- }
-
- BUG_ON(!(trans->flags & BTREE_INSERT_ATOMIC) && ret == -EINTR);
-
- return ret;
-err:
- flags = trans->flags;
+ trans->nr_updates = dst - trans->updates;
/*
* BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
* update; if we haven't done anything yet it doesn't apply
*/
- if (!trans->did_work)
- flags &= ~BTREE_INSERT_NOUNLOCK;
+ flags &= ~BTREE_INSERT_NOUNLOCK;
switch (ret) {
case BTREE_INSERT_BTREE_NODE_FULL:
* XXX:
* split -> btree node merging (of parent node) might still drop
* locks when we're not passing it BTREE_INSERT_NOUNLOCK
+ *
+ * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
+ * will inhibit merging - but we don't have a reliable way yet
+ * (do we?) of checking if we dropped locks in this path
*/
- if (!ret && !trans->did_work)
+ if (!ret)
goto retry;
#endif
* don't care if we got ENOSPC because we told split it
* couldn't block:
*/
- if (!ret || (flags & BTREE_INSERT_NOUNLOCK)) {
- trans_restart(" (split)");
+ if (!ret ||
+ ret == -EINTR ||
+ (flags & BTREE_INSERT_NOUNLOCK)) {
+ trace_trans_restart_btree_node_split(trans->ip);
ret = -EINTR;
}
break;
ret = -ENOSPC;
break;
case BTREE_INSERT_NEED_MARK_REPLICAS:
- if (flags & BTREE_INSERT_NOUNLOCK) {
- ret = -EINTR;
- goto out;
+ bch2_trans_unlock(trans);
+
+ trans_for_each_update_iter(trans, i) {
+ ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
+ if (ret)
+ return ret;
}
- bch2_btree_iter_unlock(trans->entries[0].iter);
- ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k))
- ?: -EINTR;
+ if (bch2_trans_relock(trans))
+ return 0;
+
+ trace_trans_restart_mark_replicas(trans->ip);
+ ret = -EINTR;
+ break;
+ case BTREE_INSERT_NEED_JOURNAL_RES:
+ bch2_trans_unlock(trans);
+
+ ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
+ if (ret)
+ return ret;
+
+ if (bch2_trans_relock(trans))
+ return 0;
+
+ trace_trans_restart_journal_res_get(trans->ip);
+ ret = -EINTR;
break;
default:
BUG_ON(ret >= 0);
}
if (ret == -EINTR) {
- if (flags & BTREE_INSERT_NOUNLOCK) {
- trans_restart(" (can't unlock)");
- goto out;
- }
+ int ret2 = bch2_btree_iter_traverse_all(trans);
- trans_for_each_iter(trans, i) {
- int ret2 = bch2_btree_iter_traverse(i->iter);
- if (ret2) {
- ret = ret2;
- trans_restart(" (traverse)");
- goto out;
- }
-
- BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
+ if (ret2) {
+ trace_trans_restart_traverse(trans->ip);
+ return ret2;
}
/*
* dropped locks:
*/
if (!(flags & BTREE_INSERT_ATOMIC))
- goto retry;
+ return 0;
- trans_restart(" (atomic)");
+ trace_trans_restart_atomic(trans->ip);
}
- goto out;
+ return ret;
+}
+
+/**
+ * __bch_btree_insert_at - insert keys at given iterator positions
+ *
+ * This is main entry point for btree updates.
+ *
+ * Return values:
+ * -EINTR: locking changed, this function should be called again. Only returned
+ * if passed BTREE_INSERT_ATOMIC.
+ * -EROFS: filesystem read only
+ * -EIO: journal or btree node IO error
+ */
+static int __bch2_trans_commit(struct btree_trans *trans,
+ struct btree_insert_entry **stopped_at)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i;
+ int ret;
+
+ trans_for_each_update_iter(trans, i) {
+ if (!bch2_btree_iter_upgrade(i->iter, 1)) {
+ trace_trans_restart_upgrade(trans->ip);
+ ret = -EINTR;
+ goto err;
+ }
+
+ ret = btree_iter_err(i->iter);
+ if (ret)
+ goto err;
+ }
+
+ ret = do_btree_insert_at(trans, stopped_at);
+ if (unlikely(ret))
+ goto err;
+
+ if (trans->flags & BTREE_INSERT_NOUNLOCK)
+ trans->nounlock = true;
+
+ trans_for_each_update_leaf(trans, i)
+ bch2_foreground_maybe_merge(c, i->iter, 0, trans->flags);
+
+ trans->nounlock = false;
+
+ trans_for_each_update_iter(trans, i)
+ bch2_btree_iter_downgrade(i->iter);
+err:
+ /* make sure we didn't drop or screw up locks: */
+ bch2_btree_trans_verify_locks(trans);
+
+ return ret;
}
int bch2_trans_commit(struct btree_trans *trans,
u64 *journal_seq,
unsigned flags)
{
- struct btree_insert insert = {
- .c = trans->c,
- .disk_res = disk_res,
- .journal_seq = journal_seq,
- .flags = flags,
- .nr = trans->nr_updates,
- .entries = trans->updates,
- };
+ struct bch_fs *c = trans->c;
+ struct btree_insert_entry *i;
+ unsigned orig_mem_top = trans->mem_top;
+ int ret = 0;
if (!trans->nr_updates)
- return 0;
+ goto out_noupdates;
- trans->nr_updates = 0;
+ /* for the sake of sanity: */
+ BUG_ON(trans->nr_updates > 1 && !(flags & BTREE_INSERT_ATOMIC));
- return __bch2_btree_insert_at(&insert);
-}
+ if (flags & BTREE_INSERT_GC_LOCK_HELD)
+ lockdep_assert_held(&c->gc_lock);
-int bch2_btree_delete_at(struct btree_iter *iter, unsigned flags)
-{
- struct bkey_i k;
+ if (!trans->commit_start)
+ trans->commit_start = local_clock();
- bkey_init(&k.k);
- k.k.p = iter->pos;
+ memset(&trans->journal_res, 0, sizeof(trans->journal_res));
+ memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
+ trans->disk_res = disk_res;
+ trans->journal_seq = journal_seq;
+ trans->flags = flags;
- return bch2_btree_insert_at(iter->c, NULL, NULL,
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE|flags,
- BTREE_INSERT_ENTRY(iter, &k));
-}
+ trans_for_each_update(trans, i)
+ btree_insert_entry_checks(trans, i);
+ bch2_btree_trans_verify_locks(trans);
-int bch2_btree_insert_list_at(struct btree_iter *iter,
- struct keylist *keys,
- struct disk_reservation *disk_res,
- u64 *journal_seq, unsigned flags)
-{
- BUG_ON(flags & BTREE_INSERT_ATOMIC);
- BUG_ON(bch2_keylist_empty(keys));
- bch2_verify_keylist_sorted(keys);
+ if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
+ !percpu_ref_tryget(&c->writes))) {
+ if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
+ return -EROFS;
+
+ bch2_trans_unlock(trans);
- while (!bch2_keylist_empty(keys)) {
- int ret = bch2_btree_insert_at(iter->c, disk_res,
- journal_seq, flags,
- BTREE_INSERT_ENTRY(iter, bch2_keylist_front(keys)));
+ ret = bch2_fs_read_write_early(c);
if (ret)
return ret;
- bch2_keylist_pop_front(keys);
+ percpu_ref_get(&c->writes);
+
+ if (!bch2_trans_relock(trans)) {
+ ret = -EINTR;
+ goto err;
+ }
}
+retry:
+ ret = bch2_trans_journal_preres_get(trans);
+ if (ret)
+ goto err;
- return 0;
+ ret = __bch2_trans_commit(trans, &i);
+ if (ret)
+ goto err;
+out:
+ bch2_journal_preres_put(&c->journal, &trans->journal_preres);
+
+ if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
+ percpu_ref_put(&c->writes);
+out_noupdates:
+ if (!ret && trans->commit_start) {
+ bch2_time_stats_update(&c->times[BCH_TIME_btree_update],
+ trans->commit_start);
+ trans->commit_start = 0;
+ }
+
+ BUG_ON(!(trans->flags & BTREE_INSERT_ATOMIC) && ret == -EINTR);
+
+ if (!ret) {
+ bch2_trans_unlink_iters(trans, ~trans->iters_touched|
+ trans->iters_unlink_on_commit);
+ trans->iters_touched = 0;
+ } else {
+ bch2_trans_unlink_iters(trans, trans->iters_unlink_on_commit);
+ }
+ trans->nr_updates = 0;
+ trans->mem_top = 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) {
+ /* free memory used by triggers, they'll be reexecuted: */
+ trans->mem_top = orig_mem_top;
+ goto retry;
+ }
+
+ goto out;
+}
+
+struct btree_insert_entry *bch2_trans_update(struct btree_trans *trans,
+ struct btree_insert_entry entry)
+{
+ struct btree_insert_entry *i;
+
+ BUG_ON(trans->nr_updates >= trans->nr_iters + 4);
+
+ for (i = trans->updates;
+ i < trans->updates + trans->nr_updates;
+ i++)
+ if (btree_trans_cmp(entry, *i) < 0)
+ break;
+
+ memmove(&i[1], &i[0],
+ (void *) &trans->updates[trans->nr_updates] - (void *) i);
+ trans->nr_updates++;
+ *i = entry;
+ return i;
}
/**
- * bch_btree_insert - insert keys into the extent btree
+ * bch2_btree_insert - insert keys into the extent btree
* @c: pointer to struct bch_fs
* @id: btree to insert into
* @insert_keys: list of keys to insert
struct disk_reservation *disk_res,
u64 *journal_seq, int flags)
{
- struct btree_iter iter;
+ struct btree_trans trans;
+ struct btree_iter *iter;
int ret;
- bch2_btree_iter_init(&iter, c, id, bkey_start_pos(&k->k),
- BTREE_ITER_INTENT);
- ret = bch2_btree_insert_at(c, disk_res, journal_seq, flags,
- BTREE_INSERT_ENTRY(&iter, k));
- bch2_btree_iter_unlock(&iter);
+ 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, BTREE_INSERT_ENTRY(iter, k));
+
+ ret = bch2_trans_commit(&trans, disk_res, journal_seq, flags);
+ if (ret == -EINTR)
+ goto retry;
+ bch2_trans_exit(&trans);
return ret;
}
-/*
- * bch_btree_delete_range - delete everything within a given range
- *
- * Range is a half open interval - [start, end)
- */
-int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
- struct bpos start, struct bpos end,
- u64 *journal_seq)
+int bch2_btree_delete_at_range(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bpos end,
+ u64 *journal_seq)
{
- struct btree_iter iter;
struct bkey_s_c k;
int ret = 0;
-
- bch2_btree_iter_init(&iter, c, id, start,
- BTREE_ITER_INTENT);
-
- while ((k = bch2_btree_iter_peek(&iter)).k &&
- !(ret = btree_iter_err(k)) &&
- bkey_cmp(iter.pos, end) < 0) {
- unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits);
- /* really shouldn't be using a bare, unpadded bkey_i */
+retry:
+ while ((k = bch2_btree_iter_peek(iter)).k &&
+ !(ret = bkey_err(k)) &&
+ bkey_cmp(iter->pos, end) < 0) {
struct bkey_i delete;
bkey_init(&delete.k);
* (bch2_btree_iter_peek() does guarantee that iter.pos >=
* bkey_start_pos(k.k)).
*/
- delete.k.p = iter.pos;
+ delete.k.p = iter->pos;
+
+ if (iter->flags & BTREE_ITER_IS_EXTENTS) {
+ unsigned max_sectors =
+ KEY_SIZE_MAX & (~0 << trans->c->block_bits);
- if (iter.flags & BTREE_ITER_IS_EXTENTS) {
/* create the biggest key we can */
bch2_key_resize(&delete.k, max_sectors);
bch2_cut_back(end, &delete.k);
+ bch2_extent_trim_atomic(&delete, iter);
}
- ret = bch2_btree_insert_at(c, NULL, journal_seq,
- BTREE_INSERT_NOFAIL,
- BTREE_INSERT_ENTRY(&iter, &delete));
+ bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &delete));
+ ret = bch2_trans_commit(trans, NULL, journal_seq,
+ BTREE_INSERT_ATOMIC|
+ BTREE_INSERT_NOFAIL);
if (ret)
break;
- bch2_btree_iter_cond_resched(&iter);
+ bch2_trans_cond_resched(trans);
}
- bch2_btree_iter_unlock(&iter);
+ if (ret == -EINTR) {
+ ret = 0;
+ goto retry;
+ }
+
+ 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, BTREE_INSERT_ENTRY(iter, &k));
+ return bch2_trans_commit(trans, NULL, NULL,
+ BTREE_INSERT_NOFAIL|
+ BTREE_INSERT_USE_RESERVE|flags);
+}
+
+/*
+ * bch_btree_delete_range - delete everything within a given range
+ *
+ * Range is a half open interval - [start, end)
+ */
+int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
+ 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;
}
projects / bcachefs-tools-debian / blobdiff