#include "bcachefs.h"
#include "bkey_methods.h"
+#include "bkey_buf.h"
#include "btree_cache.h"
#include "btree_iter.h"
#include "btree_key_cache.h"
#include "btree_locking.h"
#include "btree_update.h"
#include "debug.h"
+#include "error.h"
#include "extents.h"
#include "journal.h"
+#include "replicas.h"
#include <linux/prefetch.h>
#include <trace/events/bcachefs.h>
+static void btree_iter_set_search_pos(struct btree_iter *, struct bpos);
+static struct btree_iter *btree_iter_child_alloc(struct btree_iter *, unsigned long);
+static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *);
+static void btree_iter_copy(struct btree_iter *, struct btree_iter *);
+
+static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
+{
+ EBUG_ON(btree_iter_type(iter) == BTREE_ITER_NODES);
+
+ /* Are we iterating over keys in all snapshots? */
+ if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
+ p = bpos_successor(p);
+ } else {
+ p = bpos_nosnap_successor(p);
+ p.snapshot = iter->snapshot;
+ }
+
+ return p;
+}
+
+static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
+{
+ EBUG_ON(btree_iter_type(iter) == BTREE_ITER_NODES);
+
+ /* Are we iterating over keys in all snapshots? */
+ if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
+ p = bpos_predecessor(p);
+ } else {
+ p = bpos_nosnap_predecessor(p);
+ p.snapshot = iter->snapshot;
+ }
+
+ return p;
+}
+
static inline bool is_btree_node(struct btree_iter *iter, unsigned l)
{
return l < BTREE_MAX_DEPTH &&
if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
bkey_cmp(pos, POS_MAX))
- pos = bkey_successor(pos);
+ pos = bkey_successor(iter, pos);
return pos;
}
static inline bool btree_iter_pos_before_node(struct btree_iter *iter,
struct btree *b)
{
- return bkey_cmp(btree_iter_search_key(iter), b->data->min_key) < 0;
+ return bpos_cmp(iter->real_pos, b->data->min_key) < 0;
}
static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
struct btree *b)
{
- return bkey_cmp(b->key.k.p, btree_iter_search_key(iter)) < 0;
+ return bpos_cmp(b->key.k.p, iter->real_pos) < 0;
}
static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
return true;
}
-static inline bool btree_iter_get_locks(struct btree_iter *iter,
- bool upgrade, bool trace)
+static inline bool btree_iter_get_locks(struct btree_iter *iter, bool upgrade,
+ unsigned long trace_ip)
{
unsigned l = iter->level;
int fail_idx = -1;
if (!(upgrade
? bch2_btree_node_upgrade(iter, l)
: bch2_btree_node_relock(iter, l))) {
- if (trace)
- (upgrade
- ? trace_node_upgrade_fail
- : trace_node_relock_fail)(l, iter->l[l].lock_seq,
- is_btree_node(iter, l)
- ? 0
- : (unsigned long) iter->l[l].b,
- is_btree_node(iter, l)
- ? iter->l[l].b->c.lock.state.seq
- : 0);
+ (upgrade
+ ? trace_node_upgrade_fail
+ : trace_node_relock_fail)(iter->trans->ip, trace_ip,
+ iter->btree_id, &iter->real_pos,
+ l, iter->l[l].lock_seq,
+ is_btree_node(iter, l)
+ ? 0
+ : (unsigned long) iter->l[l].b,
+ is_btree_node(iter, l)
+ ? iter->l[l].b->c.lock.state.seq
+ : 0);
fail_idx = l;
btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
struct btree_iter *linked, *deadlock_iter = NULL;
u64 start_time = local_clock();
unsigned reason = 9;
+ bool ret;
/* Check if it's safe to block: */
trans_for_each_iter(trans, linked) {
*/
if (type == SIX_LOCK_intent &&
linked->nodes_locked != linked->nodes_intent_locked) {
- if (!(trans->nounlock)) {
- linked->locks_want = max_t(unsigned,
- linked->locks_want,
- __fls(linked->nodes_locked) + 1);
- if (!btree_iter_get_locks(linked, true, false)) {
- deadlock_iter = linked;
- reason = 1;
- }
- } else {
- deadlock_iter = linked;
- reason = 2;
- }
+ deadlock_iter = linked;
+ reason = 1;
}
if (linked->btree_id != iter->btree_id) {
* we're about to lock, it must have the ancestors locked too:
*/
if (level > __fls(linked->nodes_locked)) {
- if (!(trans->nounlock)) {
- linked->locks_want =
- max(level + 1, max_t(unsigned,
- linked->locks_want,
- iter->locks_want));
- if (!btree_iter_get_locks(linked, true, false)) {
- deadlock_iter = linked;
- reason = 5;
- }
- } else {
- deadlock_iter = linked;
- reason = 6;
- }
+ deadlock_iter = linked;
+ reason = 5;
}
/* Must lock btree nodes in key order: */
if (btree_node_locked(linked, level) &&
- bkey_cmp(pos, btree_node_pos((void *) linked->l[level].b,
+ bpos_cmp(pos, btree_node_pos((void *) linked->l[level].b,
btree_iter_type(linked))) <= 0) {
deadlock_iter = linked;
reason = 7;
- }
-
- /*
- * Recheck if this is a node we already have locked - since one
- * of the get_locks() calls might've successfully
- * upgraded/relocked it:
- */
- if (linked->l[level].b == b &&
- btree_node_locked_type(linked, level) >= type) {
- six_lock_increment(&b->c.lock, type);
- return true;
+ BUG_ON(trans->in_traverse_all);
}
}
if (unlikely(deadlock_iter)) {
trace_trans_restart_would_deadlock(iter->trans->ip, ip,
- reason,
+ trans->in_traverse_all, reason,
deadlock_iter->btree_id,
btree_iter_type(deadlock_iter),
+ &deadlock_iter->real_pos,
iter->btree_id,
- btree_iter_type(iter));
+ btree_iter_type(iter),
+ &pos);
return false;
}
if (six_trylock_type(&b->c.lock, type))
return true;
- if (six_lock_type(&b->c.lock, type, should_sleep_fn, p))
- return false;
+#ifdef CONFIG_BCACHEFS_DEBUG
+ trans->locking_iter_idx = iter->idx;
+ trans->locking_pos = pos;
+ trans->locking_btree_id = iter->btree_id;
+ trans->locking_level = level;
+ trans->locking = b;
+#endif
- bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)],
- start_time);
- return true;
+ ret = six_lock_type(&b->c.lock, type, should_sleep_fn, p) == 0;
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+ trans->locking = NULL;
+#endif
+ if (ret)
+ bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)],
+ start_time);
+ return ret;
}
/* Btree iterator locking: */
{
struct btree_iter *iter;
- trans_for_each_iter_all(trans, iter)
+ trans_for_each_iter(trans, iter)
bch2_btree_iter_verify_locks(iter);
}
#else
#endif
__flatten
-bool bch2_btree_iter_relock(struct btree_iter *iter, bool trace)
+static bool bch2_btree_iter_relock(struct btree_iter *iter, unsigned long trace_ip)
{
- return btree_iter_get_locks(iter, false, trace);
+ return btree_iter_get_locks(iter, false, trace_ip);
}
bool __bch2_btree_iter_upgrade(struct btree_iter *iter,
iter->locks_want = new_locks_want;
- if (btree_iter_get_locks(iter, true, true))
+ if (btree_iter_get_locks(iter, true, _THIS_IP_))
return true;
/*
- * Ancestor nodes must be locked before child nodes, so set locks_want
- * on iterators that might lock ancestors before us to avoid getting
- * -EINTR later:
+ * XXX: this is ugly - we'd prefer to not be mucking with other
+ * iterators in the btree_trans here.
+ *
+ * On failure to upgrade the iterator, setting iter->locks_want and
+ * calling get_locks() is sufficient to make bch2_btree_iter_traverse()
+ * get the locks we want on transaction restart.
+ *
+ * But if this iterator was a clone, on transaction restart what we did
+ * to this iterator isn't going to be preserved.
+ *
+ * Possibly we could add an iterator field for the parent iterator when
+ * an iterator is a copy - for now, we'll just upgrade any other
+ * iterators with the same btree id.
+ *
+ * The code below used to be needed to ensure ancestor nodes get locked
+ * before interior nodes - now that's handled by
+ * bch2_btree_iter_traverse_all().
*/
trans_for_each_iter(iter->trans, linked)
if (linked != iter &&
+ btree_iter_type(linked) == btree_iter_type(iter) &&
linked->btree_id == iter->btree_id &&
linked->locks_want < new_locks_want) {
linked->locks_want = new_locks_want;
- btree_iter_get_locks(linked, true, false);
+ btree_iter_get_locks(linked, true, _THIS_IP_);
}
return false;
}
-bool __bch2_btree_iter_upgrade_nounlock(struct btree_iter *iter,
- unsigned new_locks_want)
+void __bch2_btree_iter_downgrade(struct btree_iter *iter,
+ unsigned new_locks_want)
{
- unsigned l = iter->level;
+ unsigned l;
- EBUG_ON(iter->locks_want >= new_locks_want);
+ EBUG_ON(iter->locks_want < new_locks_want);
iter->locks_want = new_locks_want;
- do {
- if (!btree_iter_node(iter, l))
- break;
-
- if (!bch2_btree_node_upgrade(iter, l)) {
- iter->locks_want = l;
- return false;
- }
-
- l++;
- } while (l < iter->locks_want);
-
- return true;
-}
-
-void __bch2_btree_iter_downgrade(struct btree_iter *iter,
- unsigned downgrade_to)
-{
- unsigned l, new_locks_want = downgrade_to ?:
- (iter->flags & BTREE_ITER_INTENT ? 1 : 0);
-
- if (iter->locks_want < downgrade_to) {
- iter->locks_want = new_locks_want;
-
- while (iter->nodes_locked &&
- (l = __fls(iter->nodes_locked)) >= iter->locks_want) {
- if (l > iter->level) {
- btree_node_unlock(iter, l);
- } else {
- if (btree_node_intent_locked(iter, l)) {
- six_lock_downgrade(&iter->l[l].b->c.lock);
- iter->nodes_intent_locked ^= 1 << l;
- }
- break;
+ while (iter->nodes_locked &&
+ (l = __fls(iter->nodes_locked)) >= iter->locks_want) {
+ if (l > iter->level) {
+ btree_node_unlock(iter, l);
+ } else {
+ if (btree_node_intent_locked(iter, l)) {
+ six_lock_downgrade(&iter->l[l].b->c.lock);
+ iter->nodes_intent_locked ^= 1 << l;
}
+ break;
}
}
/* Btree transaction locking: */
+static inline bool btree_iter_should_be_locked(struct btree_trans *trans,
+ struct btree_iter *iter)
+{
+ return (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) ||
+ iter->should_be_locked;
+}
+
bool bch2_trans_relock(struct btree_trans *trans)
{
struct btree_iter *iter;
- bool ret = true;
trans_for_each_iter(trans, iter)
- if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
- ret &= bch2_btree_iter_relock(iter, true);
-
- return ret;
+ if (!bch2_btree_iter_relock(iter, _RET_IP_) &&
+ btree_iter_should_be_locked(trans, iter)) {
+ trace_trans_restart_relock(trans->ip, _RET_IP_,
+ iter->btree_id, &iter->real_pos);
+ return false;
+ }
+ return true;
}
void bch2_trans_unlock(struct btree_trans *trans)
static void bch2_btree_iter_verify_level(struct btree_iter *iter,
unsigned level)
{
- struct bpos pos = btree_iter_search_key(iter);
- struct btree_iter_level *l = &iter->l[level];
- struct btree_node_iter tmp = l->iter;
- bool locked = btree_node_locked(iter, level);
+ struct btree_iter_level *l;
+ struct btree_node_iter tmp;
+ bool locked;
struct bkey_packed *p, *k;
- char buf1[100], buf2[100];
+ char buf1[100], buf2[100], buf3[100];
const char *msg;
if (!bch2_debug_check_iterators)
return;
+ l = &iter->l[level];
+ tmp = l->iter;
+ locked = btree_node_locked(iter, level);
+
if (btree_iter_type(iter) == BTREE_ITER_CACHED) {
if (!level)
bch2_btree_iter_verify_cached(iter);
if (!bch2_btree_node_relock(iter, level))
return;
- /*
- * Ideally this invariant would always be true, and hopefully in the
- * future it will be, but for now set_pos_same_leaf() breaks it:
- */
- BUG_ON(iter->uptodate < BTREE_ITER_NEED_TRAVERSE &&
- !btree_iter_pos_in_node(iter, l->b));
+ BUG_ON(!btree_iter_pos_in_node(iter, l->b));
/*
* node iterators don't use leaf node iterator:
* whiteouts)
*/
p = level || btree_node_type_is_extents(iter->btree_id)
- ? bch2_btree_node_iter_prev_filter(&tmp, l->b, KEY_TYPE_discard)
+ ? bch2_btree_node_iter_prev(&tmp, l->b)
: bch2_btree_node_iter_prev_all(&tmp, l->b);
k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
- if (p && bkey_iter_pos_cmp(l->b, p, &pos) >= 0) {
+ if (p && bkey_iter_pos_cmp(l->b, p, &iter->real_pos) >= 0) {
msg = "before";
goto err;
}
- if (k && bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
+ if (k && bkey_iter_pos_cmp(l->b, k, &iter->real_pos) < 0) {
msg = "after";
goto err;
}
btree_node_unlock(iter, level);
return;
err:
- strcpy(buf1, "(none)");
strcpy(buf2, "(none)");
+ strcpy(buf3, "(none)");
+
+ bch2_bpos_to_text(&PBUF(buf1), iter->real_pos);
if (p) {
struct bkey uk = bkey_unpack_key(l->b, p);
- bch2_bkey_to_text(&PBUF(buf1), &uk);
+ bch2_bkey_to_text(&PBUF(buf2), &uk);
}
if (k) {
struct bkey uk = bkey_unpack_key(l->b, k);
- bch2_bkey_to_text(&PBUF(buf2), &uk);
+ bch2_bkey_to_text(&PBUF(buf3), &uk);
}
panic("iterator should be %s key at level %u:\n"
- "iter pos %s %llu:%llu\n"
+ "iter pos %s\n"
"prev key %s\n"
"cur key %s\n",
- msg, level,
- iter->flags & BTREE_ITER_IS_EXTENTS ? ">" : "=>",
- iter->pos.inode, iter->pos.offset,
- buf1, buf2);
+ msg, level, buf1, buf2, buf3);
}
static void bch2_btree_iter_verify(struct btree_iter *iter)
{
+ enum btree_iter_type type = btree_iter_type(iter);
unsigned i;
- bch2_btree_trans_verify_locks(iter->trans);
+ EBUG_ON(iter->btree_id >= BTREE_ID_NR);
+
+ BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
+ iter->pos.snapshot != iter->snapshot);
+
+ BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
+ (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
+
+ BUG_ON(type == BTREE_ITER_NODES &&
+ !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
+
+ BUG_ON(type != BTREE_ITER_NODES &&
+ (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
+ !btree_type_has_snapshots(iter->btree_id));
+
+ bch2_btree_iter_verify_locks(iter);
for (i = 0; i < BTREE_MAX_DEPTH; i++)
bch2_btree_iter_verify_level(iter, i);
}
+static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
+{
+ enum btree_iter_type type = btree_iter_type(iter);
+
+ BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
+ iter->pos.snapshot != iter->snapshot);
+
+ BUG_ON((type == BTREE_ITER_KEYS ||
+ type == BTREE_ITER_CACHED) &&
+ (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
+ bkey_cmp(iter->pos, iter->k.p) > 0));
+}
+
void bch2_btree_trans_verify_iters(struct btree_trans *trans, struct btree *b)
{
struct btree_iter *iter;
static inline void bch2_btree_iter_verify_level(struct btree_iter *iter, unsigned l) {}
static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
+static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
#endif
struct bkey_packed *where)
{
struct btree_iter_level *l = &iter->l[b->c.level];
- struct bpos pos = btree_iter_search_key(iter);
if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
return;
- if (bkey_iter_pos_cmp(l->b, where, &pos) < 0)
+ if (bkey_iter_pos_cmp(l->b, where, &iter->real_pos) < 0)
bch2_btree_node_iter_advance(&l->iter, l->b);
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
bool iter_current_key_modified =
orig_iter_pos >= offset &&
orig_iter_pos <= offset + clobber_u64s;
- struct bpos iter_pos = btree_iter_search_key(iter);
btree_node_iter_for_each(node_iter, set)
if (set->end == old_end)
/* didn't find the bset in the iterator - might have to readd it: */
if (new_u64s &&
- bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
+ bkey_iter_pos_cmp(b, where, &iter->real_pos) >= 0) {
bch2_btree_node_iter_push(node_iter, b, where, end);
goto fixup_done;
} else {
return;
if (new_u64s &&
- bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
+ bkey_iter_pos_cmp(b, where, &iter->real_pos) >= 0) {
set->k = offset;
} else if (set->k < offset + clobber_u64s) {
set->k = offset + new_u64s;
ret = bkey_disassemble(l->b, k, u);
- if (bch2_debug_check_bkeys)
+ /*
+ * XXX: bch2_btree_bset_insert_key() generates invalid keys when we
+ * overwrite extents - it sets k->type = KEY_TYPE_deleted on the key
+ * being overwritten but doesn't change k->size. But this is ok, because
+ * those keys are never written out, we just have to avoid a spurious
+ * assertion here:
+ */
+ if (bch2_debug_check_bkeys && !bkey_deleted(ret.k))
bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
return ret;
}
/* peek_all() doesn't skip deleted keys */
-static inline struct bkey_s_c __btree_iter_peek_all(struct btree_iter *iter,
- struct btree_iter_level *l,
- struct bkey *u)
+static inline struct bkey_s_c btree_iter_level_peek_all(struct btree_iter *iter,
+ struct btree_iter_level *l)
{
- return __btree_iter_unpack(iter, l, u,
+ return __btree_iter_unpack(iter, l, &iter->k,
bch2_btree_node_iter_peek_all(&l->iter, l->b));
}
-static inline struct bkey_s_c __btree_iter_peek(struct btree_iter *iter,
- struct btree_iter_level *l)
+static inline struct bkey_s_c btree_iter_level_peek(struct btree_iter *iter,
+ struct btree_iter_level *l)
{
- return __btree_iter_unpack(iter, l, &iter->k,
+ struct bkey_s_c k = __btree_iter_unpack(iter, l, &iter->k,
bch2_btree_node_iter_peek(&l->iter, l->b));
+
+ iter->real_pos = k.k ? k.k->p : l->b->key.k.p;
+ return k;
}
-static inline struct bkey_s_c __btree_iter_prev(struct btree_iter *iter,
- struct btree_iter_level *l)
+static inline struct bkey_s_c btree_iter_level_prev(struct btree_iter *iter,
+ struct btree_iter_level *l)
{
- return __btree_iter_unpack(iter, l, &iter->k,
+ struct bkey_s_c k = __btree_iter_unpack(iter, l, &iter->k,
bch2_btree_node_iter_prev(&l->iter, l->b));
+
+ iter->real_pos = k.k ? k.k->p : l->b->data->min_key;
+ return k;
}
static inline bool btree_iter_advance_to_pos(struct btree_iter *iter,
struct btree_iter_level *l,
int max_advance)
{
- struct bpos pos = btree_iter_search_key(iter);
struct bkey_packed *k;
int nr_advanced = 0;
while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
- bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
+ bkey_iter_pos_cmp(l->b, k, &iter->real_pos) < 0) {
if (max_advance > 0 && nr_advanced >= max_advance)
return false;
if (!k ||
bkey_deleted(k) ||
bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
- char buf[100];
+ char buf1[100];
+ char buf2[100];
+ char buf3[100];
+ char buf4[100];
struct bkey uk = bkey_unpack_key(b, k);
- bch2_bkey_to_text(&PBUF(buf), &uk);
- panic("parent iter doesn't point to new node:\n%s\n%llu:%llu\n",
- buf, b->key.k.p.inode, b->key.k.p.offset);
+ bch2_dump_btree_node(iter->trans->c, l->b);
+ bch2_bpos_to_text(&PBUF(buf1), iter->real_pos);
+ bch2_bkey_to_text(&PBUF(buf2), &uk);
+ bch2_bpos_to_text(&PBUF(buf3), b->data->min_key);
+ bch2_bpos_to_text(&PBUF(buf3), b->data->max_key);
+ panic("parent iter doesn't point to new node:\n"
+ "iter pos %s %s\n"
+ "iter key %s\n"
+ "new node %s-%s\n",
+ bch2_btree_ids[iter->btree_id], buf1,
+ buf2, buf3, buf4);
}
if (!parent_locked)
static inline void __btree_iter_init(struct btree_iter *iter,
unsigned level)
{
- struct bpos pos = btree_iter_search_key(iter);
struct btree_iter_level *l = &iter->l[level];
- bch2_btree_node_iter_init(&l->iter, l->b, &pos);
+ bch2_btree_node_iter_init(&l->iter, l->b, &iter->real_pos);
+
+ /*
+ * Iterators to interior nodes should always be pointed at the first non
+ * whiteout:
+ */
+ if (level)
+ bch2_btree_node_iter_peek(&l->iter, l->b);
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
}
trans_for_each_iter(iter->trans, linked)
if (linked->l[level].b == b) {
- __btree_node_unlock(linked, level);
+ btree_node_unlock(linked, level);
linked->l[level].b = BTREE_ITER_NO_NODE_DROP;
}
}
struct btree_iter_level *l = &iter->l[iter->level];
struct btree_node_iter node_iter = l->iter;
struct bkey_packed *k;
- BKEY_PADDED(k) tmp;
+ struct bkey_buf tmp;
unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
? (iter->level > 1 ? 0 : 2)
: (iter->level > 1 ? 1 : 16);
bool was_locked = btree_node_locked(iter, iter->level);
+ bch2_bkey_buf_init(&tmp);
+
while (nr) {
if (!bch2_btree_node_relock(iter, iter->level))
- return;
+ break;
bch2_btree_node_iter_advance(&node_iter, l->b);
k = bch2_btree_node_iter_peek(&node_iter, l->b);
if (!k)
break;
- bch2_bkey_unpack(l->b, &tmp.k, k);
- bch2_btree_node_prefetch(c, iter, &tmp.k, iter->level - 1);
+ bch2_bkey_buf_unpack(&tmp, c, l->b, k);
+ bch2_btree_node_prefetch(c, iter, tmp.k, iter->btree_id,
+ iter->level - 1);
}
if (!was_locked)
btree_node_unlock(iter, iter->level);
+
+ bch2_bkey_buf_exit(&tmp, c);
}
static noinline void btree_node_mem_ptr_set(struct btree_iter *iter,
struct btree *b;
unsigned level = iter->level - 1;
enum six_lock_type lock_type = __btree_lock_want(iter, level);
- BKEY_PADDED(k) tmp;
+ struct bkey_buf tmp;
+ int ret;
EBUG_ON(!btree_node_locked(iter, iter->level));
- bch2_bkey_unpack(l->b, &tmp.k,
+ bch2_bkey_buf_init(&tmp);
+ bch2_bkey_buf_unpack(&tmp, c, l->b,
bch2_btree_node_iter_peek(&l->iter, l->b));
- b = bch2_btree_node_get(c, iter, &tmp.k, level, lock_type, trace_ip);
- if (unlikely(IS_ERR(b)))
- return PTR_ERR(b);
+ b = bch2_btree_node_get(c, iter, tmp.k, level, lock_type, trace_ip);
+ ret = PTR_ERR_OR_ZERO(b);
+ if (unlikely(ret))
+ goto err;
mark_btree_node_locked(iter, level, lock_type);
btree_iter_node_set(iter, b);
- if (tmp.k.k.type == KEY_TYPE_btree_ptr_v2 &&
- unlikely(b != btree_node_mem_ptr(&tmp.k)))
+ if (tmp.k->k.type == KEY_TYPE_btree_ptr_v2 &&
+ unlikely(b != btree_node_mem_ptr(tmp.k)))
btree_node_mem_ptr_set(iter, level + 1, b);
if (iter->flags & BTREE_ITER_PREFETCH)
btree_iter_prefetch(iter);
+ if (btree_node_read_locked(iter, level + 1))
+ btree_node_unlock(iter, level + 1);
iter->level = level;
- return 0;
-}
-
-static void btree_iter_up(struct btree_iter *iter)
-{
- btree_node_unlock(iter, iter->level++);
+ bch2_btree_iter_verify_locks(iter);
+err:
+ bch2_bkey_buf_exit(&tmp, c);
+ return ret;
}
static int btree_iter_traverse_one(struct btree_iter *, unsigned long);
-static int __btree_iter_traverse_all(struct btree_trans *trans, int ret)
+static int __btree_iter_traverse_all(struct btree_trans *trans, int ret,
+ unsigned long trace_ip)
{
struct bch_fs *c = trans->c;
struct btree_iter *iter;
u8 sorted[BTREE_ITER_MAX];
- unsigned i, nr_sorted = 0;
+ int i, nr_sorted = 0;
+ bool relock_fail;
if (trans->in_traverse_all)
return -EINTR;
trans->in_traverse_all = true;
retry_all:
nr_sorted = 0;
+ relock_fail = false;
- trans_for_each_iter(trans, iter)
+ trans_for_each_iter(trans, iter) {
+ if (!bch2_btree_iter_relock(iter, _THIS_IP_))
+ relock_fail = true;
sorted[nr_sorted++] = iter->idx;
+ }
+
+ if (!relock_fail) {
+ trans->in_traverse_all = false;
+ return 0;
+ }
#define btree_iter_cmp_by_idx(_l, _r) \
btree_iter_lock_cmp(&trans->iters[_l], &trans->iters[_r])
bubble_sort(sorted, nr_sorted, btree_iter_cmp_by_idx);
#undef btree_iter_cmp_by_idx
+
+ for (i = nr_sorted - 2; i >= 0; --i) {
+ struct btree_iter *iter1 = trans->iters + sorted[i];
+ struct btree_iter *iter2 = trans->iters + sorted[i + 1];
+
+ if (iter1->btree_id == iter2->btree_id &&
+ iter1->locks_want < iter2->locks_want)
+ __bch2_btree_iter_upgrade(iter1, iter2->locks_want);
+ else if (!iter1->locks_want && iter2->locks_want)
+ __bch2_btree_iter_upgrade(iter1, 1);
+ }
+
bch2_trans_unlock(trans);
cond_resched();
bch2_btree_cache_cannibalize_unlock(c);
trans->in_traverse_all = false;
+
+ trace_trans_traverse_all(trans->ip, trace_ip);
return ret;
}
int bch2_btree_iter_traverse_all(struct btree_trans *trans)
{
- return __btree_iter_traverse_all(trans, 0);
+ return __btree_iter_traverse_all(trans, 0, _RET_IP_);
}
static inline bool btree_iter_good_node(struct btree_iter *iter,
!bch2_btree_node_relock(iter, l))
return false;
- if (check_pos <= 0 && btree_iter_pos_before_node(iter, iter->l[l].b))
+ if (check_pos < 0 && btree_iter_pos_before_node(iter, iter->l[l].b))
return false;
- if (check_pos >= 0 && btree_iter_pos_after_node(iter, iter->l[l].b))
+ if (check_pos > 0 && btree_iter_pos_after_node(iter, iter->l[l].b))
return false;
return true;
}
unsigned long trace_ip)
{
unsigned depth_want = iter->level;
+ int ret = 0;
/*
* if we need interior nodes locked, call btree_iter_relock() to make
*/
if (iter->uptodate == BTREE_ITER_NEED_RELOCK ||
iter->locks_want > 1)
- bch2_btree_iter_relock(iter, false);
+ bch2_btree_iter_relock(iter, _THIS_IP_);
- if (btree_iter_type(iter) == BTREE_ITER_CACHED)
- return bch2_btree_iter_traverse_cached(iter);
+ if (btree_iter_type(iter) == BTREE_ITER_CACHED) {
+ ret = bch2_btree_iter_traverse_cached(iter);
+ goto out;
+ }
if (iter->uptodate < BTREE_ITER_NEED_RELOCK)
- return 0;
+ goto out;
if (unlikely(iter->level >= BTREE_MAX_DEPTH))
- return 0;
+ goto out;
- /*
- * XXX: correctly using BTREE_ITER_UPTODATE should make using check_pos
- * here unnecessary
- */
iter->level = btree_iter_up_until_good_node(iter, 0);
- /*
- * If we've got a btree node locked (i.e. we aren't about to relock the
- * root) - advance its node iterator if necessary:
- *
- * XXX correctly using BTREE_ITER_UPTODATE should make this unnecessary
- */
- if (is_btree_node(iter, iter->level)) {
- BUG_ON(!btree_iter_pos_in_node(iter, iter->l[iter->level].b));
-
- btree_iter_advance_to_pos(iter, &iter->l[iter->level], -1);
- }
-
/*
* Note: iter->nodes[iter->level] may be temporarily NULL here - that
* would indicate to other code that we got to the end of the btree,
* btree_iter_lock_root() comes next and that it can't fail
*/
while (iter->level > depth_want) {
- int ret = btree_iter_node(iter, iter->level)
+ ret = btree_iter_node(iter, iter->level)
? btree_iter_down(iter, trace_ip)
: btree_iter_lock_root(iter, depth_want, trace_ip);
if (unlikely(ret)) {
- if (ret == 1)
- return 0;
+ if (ret == 1) {
+ /*
+ * Got to the end of the btree (in
+ * BTREE_ITER_NODES mode)
+ */
+ ret = 0;
+ goto out;
+ }
iter->level = depth_want;
iter->l[iter->level].b =
BTREE_ITER_NO_NODE_DOWN;
}
- return ret;
+ goto out;
}
}
iter->uptodate = BTREE_ITER_NEED_PEEK;
-
+out:
+ trace_iter_traverse(iter->trans->ip, trace_ip,
+ iter->btree_id, &iter->real_pos, ret);
bch2_btree_iter_verify(iter);
- return 0;
+ return ret;
}
-int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
+static int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
{
struct btree_trans *trans = iter->trans;
int ret;
ret = bch2_trans_cond_resched(trans) ?:
btree_iter_traverse_one(iter, _RET_IP_);
if (unlikely(ret))
- ret = __btree_iter_traverse_all(trans, ret);
+ ret = __btree_iter_traverse_all(trans, ret, _RET_IP_);
return ret;
}
-static inline void bch2_btree_iter_checks(struct btree_iter *iter)
+/*
+ * Note:
+ * bch2_btree_iter_traverse() is for external users, btree_iter_traverse() is
+ * for internal btree iterator users
+ *
+ * bch2_btree_iter_traverse sets iter->real_pos to iter->pos,
+ * btree_iter_traverse() does not:
+ */
+static inline int __must_check
+btree_iter_traverse(struct btree_iter *iter)
{
- enum btree_iter_type type = btree_iter_type(iter);
+ return iter->uptodate >= BTREE_ITER_NEED_RELOCK
+ ? __bch2_btree_iter_traverse(iter)
+ : 0;
+}
- EBUG_ON(iter->btree_id >= BTREE_ID_NR);
+int __must_check
+bch2_btree_iter_traverse(struct btree_iter *iter)
+{
+ int ret;
- BUG_ON((type == BTREE_ITER_KEYS ||
- type == BTREE_ITER_CACHED) &&
- (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
- bkey_cmp(iter->pos, iter->k.p) > 0));
+ btree_iter_set_search_pos(iter, btree_iter_search_key(iter));
- bch2_btree_iter_verify_locks(iter);
- bch2_btree_iter_verify_level(iter, iter->level);
+ ret = btree_iter_traverse(iter);
+ if (ret)
+ return ret;
+
+ iter->should_be_locked = true;
+ return 0;
}
/* Iterate across nodes (leaf and interior nodes) */
int ret;
EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
- bch2_btree_iter_checks(iter);
-
- if (iter->uptodate == BTREE_ITER_UPTODATE)
- return iter->l[iter->level].b;
+ bch2_btree_iter_verify(iter);
- ret = bch2_btree_iter_traverse(iter);
+ ret = btree_iter_traverse(iter);
if (ret)
return NULL;
if (!b)
return NULL;
- BUG_ON(bkey_cmp(b->key.k.p, iter->pos) < 0);
+ BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0);
- iter->pos = b->key.k.p;
- iter->uptodate = BTREE_ITER_UPTODATE;
+ iter->pos = iter->real_pos = b->key.k.p;
bch2_btree_iter_verify(iter);
+ iter->should_be_locked = true;
return b;
}
int ret;
EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
- bch2_btree_iter_checks(iter);
+ bch2_btree_iter_verify(iter);
/* already got to end? */
if (!btree_iter_node(iter, iter->level))
bch2_trans_cond_resched(iter->trans);
- btree_iter_up(iter);
-
- if (!bch2_btree_node_relock(iter, iter->level))
- btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
+ btree_node_unlock(iter, iter->level);
+ iter->l[iter->level].b = BTREE_ITER_NO_NODE_UP;
+ iter->level++;
- ret = bch2_btree_iter_traverse(iter);
+ btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
+ ret = btree_iter_traverse(iter);
if (ret)
return NULL;
if (!b)
return NULL;
- if (bkey_cmp(iter->pos, b->key.k.p) < 0) {
+ if (bpos_cmp(iter->pos, b->key.k.p) < 0) {
/*
* Haven't gotten to the end of the parent node: go back down to
* the next child node
*/
+ btree_iter_set_search_pos(iter, bpos_successor(iter->pos));
- /*
- * We don't really want to be unlocking here except we can't
- * directly tell btree_iter_traverse() "traverse to this level"
- * except by setting iter->level, so we have to unlock so we
- * don't screw up our lock invariants:
- */
- if (btree_node_read_locked(iter, iter->level))
- btree_node_unlock(iter, iter->level);
-
- iter->pos = bkey_successor(iter->pos);
- iter->level = iter->min_depth;
+ /* Unlock to avoid screwing up our lock invariants: */
+ btree_node_unlock(iter, iter->level);
+ iter->level = iter->min_depth;
btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
- ret = bch2_btree_iter_traverse(iter);
+ bch2_btree_iter_verify(iter);
+
+ ret = btree_iter_traverse(iter);
if (ret)
return NULL;
b = iter->l[iter->level].b;
}
- iter->pos = b->key.k.p;
- iter->uptodate = BTREE_ITER_UPTODATE;
+ iter->pos = iter->real_pos = b->key.k.p;
bch2_btree_iter_verify(iter);
+ iter->should_be_locked = true;
return b;
}
/* Iterate across keys (in leaf nodes only) */
-void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_pos)
-{
- struct btree_iter_level *l = &iter->l[0];
-
- EBUG_ON(iter->level != 0);
- EBUG_ON(bkey_cmp(new_pos, iter->pos) < 0);
- EBUG_ON(!btree_node_locked(iter, 0));
- EBUG_ON(bkey_cmp(new_pos, l->b->key.k.p) > 0);
-
- bkey_init(&iter->k);
- iter->k.p = iter->pos = new_pos;
- btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
-
- btree_iter_advance_to_pos(iter, l, -1);
-
- /*
- * XXX:
- * keeping a node locked that's outside (even just outside) iter->pos
- * breaks __bch2_btree_node_lock(). This seems to only affect
- * bch2_btree_node_get_sibling so for now it's fixed there, but we
- * should try to get rid of this corner case.
- *
- * (this behaviour is currently needed for BTREE_INSERT_NOUNLOCK)
- */
-
- if (bch2_btree_node_iter_end(&l->iter) &&
- btree_iter_pos_after_node(iter, l->b))
- btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
-}
-
-static void btree_iter_pos_changed(struct btree_iter *iter, int cmp)
+static void btree_iter_set_search_pos(struct btree_iter *iter, struct bpos new_pos)
{
+#ifdef CONFIG_BCACHEFS_DEBUG
+ struct bpos old_pos = iter->real_pos;
+#endif
+ int cmp = bpos_cmp(new_pos, iter->real_pos);
unsigned l = iter->level;
if (!cmp)
goto out;
+ iter->real_pos = new_pos;
+ iter->should_be_locked = false;
+
if (unlikely(btree_iter_type(iter) == BTREE_ITER_CACHED)) {
btree_node_unlock(iter, 0);
- iter->l[0].b = BTREE_ITER_NO_NODE_UP;
+ iter->l[0].b = BTREE_ITER_NO_NODE_CACHED;
btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
return;
}
btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
else
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
+
+ bch2_btree_iter_verify(iter);
+#ifdef CONFIG_BCACHEFS_DEBUG
+ trace_iter_set_search_pos(iter->trans->ip, _RET_IP_,
+ iter->btree_id,
+ &old_pos, &new_pos, l);
+#endif
}
-void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos,
- bool strictly_greater)
+inline bool bch2_btree_iter_advance(struct btree_iter *iter)
{
- struct bpos old = btree_iter_search_key(iter);
- int cmp;
-
- iter->flags &= ~BTREE_ITER_IS_EXTENTS;
- iter->flags |= strictly_greater ? BTREE_ITER_IS_EXTENTS : 0;
-
- bkey_init(&iter->k);
- iter->k.p = iter->pos = new_pos;
+ struct bpos pos = iter->k.p;
+ bool ret = bpos_cmp(pos, POS_MAX) != 0;
- cmp = bkey_cmp(btree_iter_search_key(iter), old);
-
- btree_iter_pos_changed(iter, cmp);
+ if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
+ pos = bkey_successor(iter, pos);
+ bch2_btree_iter_set_pos(iter, pos);
+ return ret;
}
-void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
+inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
{
- int cmp = bkey_cmp(new_pos, iter->pos);
-
- bkey_init(&iter->k);
- iter->k.p = iter->pos = new_pos;
+ struct bpos pos = bkey_start_pos(&iter->k);
+ bool ret = bpos_cmp(pos, POS_MIN) != 0;
- btree_iter_pos_changed(iter, cmp);
+ if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
+ pos = bkey_predecessor(iter, pos);
+ bch2_btree_iter_set_pos(iter, pos);
+ return ret;
}
static inline bool btree_iter_set_pos_to_next_leaf(struct btree_iter *iter)
{
- struct btree_iter_level *l = &iter->l[0];
- bool ret;
-
- bkey_init(&iter->k);
- iter->k.p = iter->pos = l->b->key.k.p;
+ struct bpos next_pos = iter->l[0].b->key.k.p;
+ bool ret = bpos_cmp(next_pos, POS_MAX) != 0;
- ret = bkey_cmp(iter->pos, POS_MAX) != 0;
- if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
- iter->k.p = iter->pos = bkey_successor(iter->pos);
+ /*
+ * Typically, we don't want to modify iter->pos here, since that
+ * indicates where we searched from - unless we got to the end of the
+ * btree, in that case we want iter->pos to reflect that:
+ */
+ if (ret)
+ btree_iter_set_search_pos(iter, bpos_successor(next_pos));
+ else
+ bch2_btree_iter_set_pos(iter, POS_MAX);
- btree_iter_pos_changed(iter, 1);
return ret;
}
static inline bool btree_iter_set_pos_to_prev_leaf(struct btree_iter *iter)
{
- struct btree_iter_level *l = &iter->l[0];
- bool ret;
-
- bkey_init(&iter->k);
- iter->k.p = iter->pos = l->b->data->min_key;
- iter->uptodate = BTREE_ITER_NEED_TRAVERSE;
-
- ret = bkey_cmp(iter->pos, POS_MIN) != 0;
- if (ret) {
- iter->k.p = iter->pos = bkey_predecessor(iter->pos);
+ struct bpos next_pos = iter->l[0].b->data->min_key;
+ bool ret = bpos_cmp(next_pos, POS_MIN) != 0;
- if (iter->flags & BTREE_ITER_IS_EXTENTS)
- iter->k.p = iter->pos = bkey_predecessor(iter->pos);
- }
+ if (ret)
+ btree_iter_set_search_pos(iter, bpos_predecessor(next_pos));
+ else
+ bch2_btree_iter_set_pos(iter, POS_MIN);
- btree_iter_pos_changed(iter, -1);
return ret;
}
-/**
- * btree_iter_peek_uptodate - given an iterator that is uptodate, return the key
- * it currently points to
- */
-static inline struct bkey_s_c btree_iter_peek_uptodate(struct btree_iter *iter)
+static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter,
+ struct bpos pos)
{
- struct btree_iter_level *l = &iter->l[0];
- struct bkey_s_c ret = { .k = &iter->k };
-
- if (!bkey_deleted(&iter->k)) {
- struct bkey_packed *_k =
- __bch2_btree_node_iter_peek_all(&l->iter, l->b);
-
- ret.v = bkeyp_val(&l->b->format, _k);
+ struct btree_insert_entry *i;
- if (bch2_debug_check_iterators) {
- struct bkey k = bkey_unpack_key(l->b, _k);
+ if (!(iter->flags & BTREE_ITER_WITH_UPDATES))
+ return NULL;
- BUG_ON(memcmp(&k, &iter->k, sizeof(k)));
+ trans_for_each_update(iter->trans, i)
+ if ((cmp_int(iter->btree_id, i->iter->btree_id) ?:
+ bkey_cmp(pos, i->k->k.p)) <= 0) {
+ if (iter->btree_id == i->iter->btree_id)
+ return i->k;
+ break;
}
- if (bch2_debug_check_bkeys)
- bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
- }
-
- return ret;
+ return NULL;
}
/**
*/
struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
{
- struct btree_iter_level *l = &iter->l[0];
+ struct bpos search_key = btree_iter_search_key(iter);
+ struct bkey_i *next_update;
struct bkey_s_c k;
int ret;
EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
- bch2_btree_iter_checks(iter);
-
- if (iter->uptodate == BTREE_ITER_UPTODATE &&
- !bkey_deleted(&iter->k))
- return btree_iter_peek_uptodate(iter);
+ bch2_btree_iter_verify(iter);
+ bch2_btree_iter_verify_entry_exit(iter);
+start:
+ next_update = btree_trans_peek_updates(iter, search_key);
+ btree_iter_set_search_pos(iter, search_key);
while (1) {
- ret = bch2_btree_iter_traverse(iter);
+ ret = btree_iter_traverse(iter);
if (unlikely(ret))
return bkey_s_c_err(ret);
- k = __btree_iter_peek(iter, l);
- if (likely(k.k))
+ k = btree_iter_level_peek(iter, &iter->l[0]);
+
+ if (next_update &&
+ bpos_cmp(next_update->k.p, iter->real_pos) <= 0) {
+ iter->k = next_update->k;
+ k = bkey_i_to_s_c(next_update);
+ }
+
+ if (likely(k.k)) {
+ if (bkey_deleted(k.k)) {
+ search_key = bkey_successor(iter, k.k->p);
+ goto start;
+ }
+
break;
+ }
if (!btree_iter_set_pos_to_next_leaf(iter))
return bkey_s_c_null;
}
/*
- * iter->pos should always be equal to the key we just
- * returned - except extents can straddle iter->pos:
+ * iter->pos should be mononotically increasing, and always be equal to
+ * the key we just returned - except extents can straddle iter->pos:
*/
- if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
- bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
+ if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
+ iter->pos = k.k->p;
+ else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
iter->pos = bkey_start_pos(k.k);
- iter->uptodate = BTREE_ITER_UPTODATE;
-
- bch2_btree_iter_verify_level(iter, 0);
+ bch2_btree_iter_verify_entry_exit(iter);
+ bch2_btree_iter_verify(iter);
+ iter->should_be_locked = true;
return k;
}
*/
struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
{
- if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
+ if (!bch2_btree_iter_advance(iter))
return bkey_s_c_null;
- bch2_btree_iter_set_pos(iter,
- (iter->flags & BTREE_ITER_IS_EXTENTS)
- ? iter->k.p
- : bkey_successor(iter->k.p));
-
return bch2_btree_iter_peek(iter);
}
-static struct bkey_s_c __btree_trans_updates_peek(struct btree_iter *iter)
-{
- struct bpos pos = btree_iter_search_key(iter);
- struct btree_trans *trans = iter->trans;
- struct btree_insert_entry *i;
-
- trans_for_each_update2(trans, i)
- if ((cmp_int(iter->btree_id, i->iter->btree_id) ?:
- bkey_cmp(pos, i->k->k.p)) <= 0)
- break;
-
- return i < trans->updates2 + trans->nr_updates2 &&
- iter->btree_id == i->iter->btree_id
- ? bkey_i_to_s_c(i->k)
- : bkey_s_c_null;
-}
-
-static struct bkey_s_c __bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
-{
- struct btree_iter_level *l = &iter->l[0];
- struct bkey_s_c k = __btree_iter_peek(iter, l);
- struct bkey_s_c u = __btree_trans_updates_peek(iter);
-
- if (k.k && (!u.k || bkey_cmp(k.k->p, u.k->p) < 0))
- return k;
- if (u.k && bkey_cmp(u.k->p, l->b->key.k.p) <= 0) {
- iter->k = *u.k;
- return u;
- }
- return bkey_s_c_null;
-}
-
-struct bkey_s_c bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
-{
- struct bkey_s_c k;
- int ret;
-
- EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
- bch2_btree_iter_checks(iter);
-
- while (1) {
- ret = bch2_btree_iter_traverse(iter);
- if (unlikely(ret))
- return bkey_s_c_err(ret);
-
- k = __bch2_btree_iter_peek_with_updates(iter);
-
- if (k.k && bkey_deleted(k.k)) {
- bch2_btree_iter_set_pos(iter,
- (iter->flags & BTREE_ITER_IS_EXTENTS)
- ? iter->k.p
- : bkey_successor(iter->k.p));
- continue;
- }
-
- if (likely(k.k))
- break;
-
- if (!btree_iter_set_pos_to_next_leaf(iter))
- return bkey_s_c_null;
- }
-
- /*
- * iter->pos should always be equal to the key we just
- * returned - except extents can straddle iter->pos:
- */
- if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
- bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
- iter->pos = bkey_start_pos(k.k);
-
- iter->uptodate = BTREE_ITER_UPTODATE;
- return k;
-}
-
-struct bkey_s_c bch2_btree_iter_next_with_updates(struct btree_iter *iter)
-{
- if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
- return bkey_s_c_null;
-
- bch2_btree_iter_set_pos(iter,
- (iter->flags & BTREE_ITER_IS_EXTENTS)
- ? iter->k.p
- : bkey_successor(iter->k.p));
-
- return bch2_btree_iter_peek_with_updates(iter);
-}
-
/**
* bch2_btree_iter_peek_prev: returns first key less than or equal to
* iterator's current position
*/
struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
{
- struct bpos pos = iter->pos;
struct btree_iter_level *l = &iter->l[0];
struct bkey_s_c k;
int ret;
EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
- bch2_btree_iter_checks(iter);
+ EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
+ bch2_btree_iter_verify(iter);
+ bch2_btree_iter_verify_entry_exit(iter);
- if (iter->uptodate == BTREE_ITER_UPTODATE &&
- !bkey_deleted(&iter->k))
- return btree_iter_peek_uptodate(iter);
+ btree_iter_set_search_pos(iter, iter->pos);
while (1) {
- ret = bch2_btree_iter_traverse(iter);
- if (unlikely(ret))
- return bkey_s_c_err(ret);
+ ret = btree_iter_traverse(iter);
+ if (unlikely(ret)) {
+ k = bkey_s_c_err(ret);
+ goto no_key;
+ }
- k = __btree_iter_peek(iter, l);
- if (!k.k || bkey_cmp(bkey_start_pos(k.k), pos) > 0)
- k = __btree_iter_prev(iter, l);
+ k = btree_iter_level_peek(iter, l);
+ if (!k.k ||
+ ((iter->flags & BTREE_ITER_IS_EXTENTS)
+ ? bkey_cmp(bkey_start_pos(k.k), iter->pos) >= 0
+ : bkey_cmp(k.k->p, iter->pos) > 0))
+ k = btree_iter_level_prev(iter, l);
if (likely(k.k))
break;
- if (!btree_iter_set_pos_to_prev_leaf(iter))
- return bkey_s_c_null;
+ if (!btree_iter_set_pos_to_prev_leaf(iter)) {
+ k = bkey_s_c_null;
+ goto no_key;
+ }
}
- EBUG_ON(bkey_cmp(bkey_start_pos(k.k), pos) > 0);
- iter->pos = bkey_start_pos(k.k);
- iter->uptodate = BTREE_ITER_UPTODATE;
+ EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
+
+ /* Extents can straddle iter->pos: */
+ if (bkey_cmp(k.k->p, iter->pos) < 0)
+ iter->pos = k.k->p;
+out:
+ bch2_btree_iter_verify_entry_exit(iter);
+ bch2_btree_iter_verify(iter);
+ iter->should_be_locked = true;
return k;
+no_key:
+ /*
+ * btree_iter_level_peek() may have set iter->k to a key we didn't want, and
+ * then we errored going to the previous leaf - make sure it's
+ * consistent with iter->pos:
+ */
+ bkey_init(&iter->k);
+ iter->k.p = iter->pos;
+ goto out;
}
/**
*/
struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
{
- struct bpos pos = bkey_start_pos(&iter->k);
-
- EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
- bch2_btree_iter_checks(iter);
-
- if (unlikely(!bkey_cmp(pos, POS_MIN)))
+ if (!bch2_btree_iter_rewind(iter))
return bkey_s_c_null;
- bch2_btree_iter_set_pos(iter, bkey_predecessor(pos));
-
return bch2_btree_iter_peek_prev(iter);
}
-static inline struct bkey_s_c
-__bch2_btree_iter_peek_slot_extents(struct btree_iter *iter)
+struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
{
- struct btree_iter_level *l = &iter->l[0];
- struct btree_node_iter node_iter;
+ struct bpos search_key;
struct bkey_s_c k;
- struct bkey n;
int ret;
- /* keys & holes can't span inode numbers: */
- if (iter->pos.offset == KEY_OFFSET_MAX) {
+ EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS &&
+ btree_iter_type(iter) != BTREE_ITER_CACHED);
+ bch2_btree_iter_verify(iter);
+ bch2_btree_iter_verify_entry_exit(iter);
+
+ /* extents can't span inode numbers: */
+ if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
+ unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
if (iter->pos.inode == KEY_INODE_MAX)
return bkey_s_c_null;
- bch2_btree_iter_set_pos(iter, bkey_successor(iter->pos));
-
- ret = bch2_btree_iter_traverse(iter);
- if (unlikely(ret))
- return bkey_s_c_err(ret);
- }
-
- /*
- * iterator is now at the correct position for inserting at iter->pos,
- * but we need to keep iterating until we find the first non whiteout so
- * we know how big a hole we have, if any:
- */
-
- node_iter = l->iter;
- k = __btree_iter_unpack(iter, l, &iter->k,
- bch2_btree_node_iter_peek(&node_iter, l->b));
-
- if (k.k && bkey_cmp(bkey_start_pos(k.k), iter->pos) <= 0) {
- /*
- * We're not setting iter->uptodate because the node iterator
- * doesn't necessarily point at the key we're returning:
- */
-
- EBUG_ON(bkey_cmp(k.k->p, iter->pos) <= 0);
- bch2_btree_iter_verify_level(iter, 0);
- return k;
+ bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
}
- /* hole */
-
- if (!k.k)
- k.k = &l->b->key.k;
+ search_key = btree_iter_search_key(iter);
+ btree_iter_set_search_pos(iter, search_key);
- bkey_init(&n);
- n.p = iter->pos;
- bch2_key_resize(&n,
- min_t(u64, KEY_SIZE_MAX,
- (k.k->p.inode == n.p.inode
- ? bkey_start_offset(k.k)
- : KEY_OFFSET_MAX) -
- n.p.offset));
-
- EBUG_ON(!n.size);
-
- iter->k = n;
- iter->uptodate = BTREE_ITER_UPTODATE;
+ ret = btree_iter_traverse(iter);
+ if (unlikely(ret))
+ return bkey_s_c_err(ret);
- bch2_btree_iter_verify_level(iter, 0);
- return (struct bkey_s_c) { &iter->k, NULL };
-}
+ if (btree_iter_type(iter) == BTREE_ITER_CACHED ||
+ !(iter->flags & BTREE_ITER_IS_EXTENTS)) {
+ struct bkey_i *next_update;
+ struct bkey_cached *ck;
-struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
-{
- struct btree_iter_level *l = &iter->l[0];
- struct bkey_s_c k;
- int ret;
+ switch (btree_iter_type(iter)) {
+ case BTREE_ITER_KEYS:
+ k = btree_iter_level_peek_all(iter, &iter->l[0]);
+ EBUG_ON(k.k && bkey_deleted(k.k) && bkey_cmp(k.k->p, iter->pos) == 0);
+ break;
+ case BTREE_ITER_CACHED:
+ ck = (void *) iter->l[0].b;
+ EBUG_ON(iter->btree_id != ck->key.btree_id ||
+ bkey_cmp(iter->pos, ck->key.pos));
+ BUG_ON(!ck->valid);
- EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
- bch2_btree_iter_checks(iter);
+ k = bkey_i_to_s_c(ck->k);
+ break;
+ case BTREE_ITER_NODES:
+ BUG();
+ }
- if (iter->uptodate == BTREE_ITER_UPTODATE)
- return btree_iter_peek_uptodate(iter);
+ next_update = btree_trans_peek_updates(iter, search_key);
+ if (next_update &&
+ (!k.k || bpos_cmp(next_update->k.p, k.k->p) <= 0)) {
+ iter->k = next_update->k;
+ k = bkey_i_to_s_c(next_update);
+ }
+ } else {
+ if ((iter->flags & BTREE_ITER_INTENT)) {
+ struct btree_iter *child =
+ btree_iter_child_alloc(iter, _THIS_IP_);
- ret = bch2_btree_iter_traverse(iter);
- if (unlikely(ret))
- return bkey_s_c_err(ret);
+ btree_iter_copy(child, iter);
+ k = bch2_btree_iter_peek(child);
- if (iter->flags & BTREE_ITER_IS_EXTENTS)
- return __bch2_btree_iter_peek_slot_extents(iter);
+ if (k.k && !bkey_err(k))
+ iter->k = child->k;
+ } else {
+ struct bpos pos = iter->pos;
- k = __btree_iter_peek_all(iter, l, &iter->k);
+ k = bch2_btree_iter_peek(iter);
+ iter->pos = pos;
+ }
- EBUG_ON(k.k && bkey_deleted(k.k) && bkey_cmp(k.k->p, iter->pos) == 0);
+ if (unlikely(bkey_err(k)))
+ return k;
+ }
- if (!k.k || bkey_cmp(iter->pos, k.k->p)) {
- /* hole */
- bkey_init(&iter->k);
- iter->k.p = iter->pos;
- k = (struct bkey_s_c) { &iter->k, NULL };
+ if (!(iter->flags & BTREE_ITER_IS_EXTENTS)) {
+ if (!k.k ||
+ ((iter->flags & BTREE_ITER_ALL_SNAPSHOTS)
+ ? bpos_cmp(iter->pos, k.k->p)
+ : bkey_cmp(iter->pos, k.k->p))) {
+ bkey_init(&iter->k);
+ iter->k.p = iter->pos;
+ k = (struct bkey_s_c) { &iter->k, NULL };
+ }
+ } else {
+ struct bpos next = k.k ? bkey_start_pos(k.k) : POS_MAX;
+
+ if (bkey_cmp(iter->pos, next) < 0) {
+ bkey_init(&iter->k);
+ iter->k.p = iter->pos;
+ bch2_key_resize(&iter->k,
+ min_t(u64, KEY_SIZE_MAX,
+ (next.inode == iter->pos.inode
+ ? next.offset
+ : KEY_OFFSET_MAX) -
+ iter->pos.offset));
+
+ k = (struct bkey_s_c) { &iter->k, NULL };
+ EBUG_ON(!k.k->size);
+ }
}
- iter->uptodate = BTREE_ITER_UPTODATE;
- bch2_btree_iter_verify_level(iter, 0);
+ bch2_btree_iter_verify_entry_exit(iter);
+ bch2_btree_iter_verify(iter);
+ iter->should_be_locked = true;
+
return k;
}
struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
{
- if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
+ if (!bch2_btree_iter_advance(iter))
return bkey_s_c_null;
- bch2_btree_iter_set_pos(iter,
- (iter->flags & BTREE_ITER_IS_EXTENTS)
- ? iter->k.p
- : bkey_successor(iter->k.p));
-
return bch2_btree_iter_peek_slot(iter);
}
-struct bkey_s_c bch2_btree_iter_peek_cached(struct btree_iter *iter)
+struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
{
- struct bkey_cached *ck;
- int ret;
-
- EBUG_ON(btree_iter_type(iter) != BTREE_ITER_CACHED);
- bch2_btree_iter_checks(iter);
-
- ret = bch2_btree_iter_traverse(iter);
- if (unlikely(ret))
- return bkey_s_c_err(ret);
-
- ck = (void *) iter->l[0].b;
-
- EBUG_ON(iter->btree_id != ck->key.btree_id ||
- bkey_cmp(iter->pos, ck->key.pos));
- BUG_ON(!ck->valid);
+ if (!bch2_btree_iter_rewind(iter))
+ return bkey_s_c_null;
- return bkey_i_to_s_c(ck->k);
+ return bch2_btree_iter_peek_slot(iter);
}
static inline void bch2_btree_iter_init(struct btree_trans *trans,
- struct btree_iter *iter, enum btree_id btree_id,
- struct bpos pos, unsigned flags)
+ struct btree_iter *iter, enum btree_id btree_id)
{
struct bch_fs *c = trans->c;
unsigned i;
- if (btree_node_type_is_extents(btree_id) &&
- !(flags & BTREE_ITER_NODES))
- flags |= BTREE_ITER_IS_EXTENTS;
-
iter->trans = trans;
- iter->pos = pos;
- bkey_init(&iter->k);
- iter->k.p = pos;
- iter->flags = flags;
iter->uptodate = BTREE_ITER_NEED_TRAVERSE;
iter->btree_id = btree_id;
+ iter->real_pos = POS_MIN;
iter->level = 0;
iter->min_depth = 0;
- iter->locks_want = flags & BTREE_ITER_INTENT ? 1 : 0;
+ iter->locks_want = 0;
iter->nodes_locked = 0;
iter->nodes_intent_locked = 0;
for (i = 0; i < ARRAY_SIZE(iter->l); i++)
/* new transactional stuff: */
+static void btree_iter_child_free(struct btree_iter *iter)
+{
+ struct btree_iter *child = btree_iter_child(iter);
+
+ if (child) {
+ bch2_trans_iter_free(iter->trans, child);
+ iter->child_idx = U8_MAX;
+ }
+}
+
+static struct btree_iter *btree_iter_child_alloc(struct btree_iter *iter,
+ unsigned long ip)
+{
+ struct btree_trans *trans = iter->trans;
+ struct btree_iter *child = btree_iter_child(iter);
+
+ if (!child) {
+ child = btree_trans_iter_alloc(trans);
+ child->ip_allocated = ip;
+ iter->child_idx = child->idx;
+
+ trans->iters_live |= 1ULL << child->idx;
+ trans->iters_touched |= 1ULL << child->idx;
+ }
+
+ return child;
+}
+
static inline void __bch2_trans_iter_free(struct btree_trans *trans,
unsigned idx)
{
+ btree_iter_child_free(&trans->iters[idx]);
+
__bch2_btree_iter_unlock(&trans->iters[idx]);
trans->iters_linked &= ~(1ULL << idx);
trans->iters_live &= ~(1ULL << idx);
return 0;
BUG_ON(trans->iters + iter->idx != iter);
+ BUG_ON(!btree_iter_live(trans, iter));
ret = btree_iter_err(iter);
if (IS_ERR_OR_NULL(iter))
return 0;
- trans->iters_touched &= ~(1ULL << iter->idx);
+ set_btree_iter_dontneed(trans, iter);
return bch2_trans_iter_put(trans, iter);
}
-#if 0
-static int bch2_trans_realloc_iters(struct btree_trans *trans,
- unsigned new_size)
+noinline __cold
+static void btree_trans_iter_alloc_fail(struct btree_trans *trans)
{
- void *p, *new_iters, *new_updates, *new_updates2;
- size_t iters_bytes;
- size_t updates_bytes;
-
- new_size = roundup_pow_of_two(new_size);
-
- BUG_ON(new_size > BTREE_ITER_MAX);
-
- if (new_size <= trans->size)
- return 0;
-
- BUG_ON(trans->used_mempool);
-
- bch2_trans_unlock(trans);
-
- iters_bytes = sizeof(struct btree_iter) * new_size;
- updates_bytes = sizeof(struct btree_insert_entry) * new_size;
-
- p = kmalloc(iters_bytes +
- updates_bytes +
- updates_bytes, GFP_NOFS);
- if (p)
- goto success;
-
- p = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
- new_size = BTREE_ITER_MAX;
- trans->used_mempool = true;
-success:
- new_iters = p; p += iters_bytes;
- new_updates = p; p += updates_bytes;
- new_updates2 = p; p += updates_bytes;
-
- memcpy(new_iters, trans->iters,
- sizeof(struct btree_iter) * trans->nr_iters);
- memcpy(new_updates, trans->updates,
- sizeof(struct btree_insert_entry) * trans->nr_updates);
- memcpy(new_updates2, trans->updates2,
- sizeof(struct btree_insert_entry) * trans->nr_updates2);
-
- if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
- memset(trans->iters, POISON_FREE,
- sizeof(struct btree_iter) * trans->nr_iters +
- sizeof(struct btree_insert_entry) * trans->nr_iters);
-
- kfree(trans->iters);
-
- trans->iters = new_iters;
- trans->updates = new_updates;
- trans->updates2 = new_updates2;
- trans->size = new_size;
+ struct btree_iter *iter;
+ struct btree_insert_entry *i;
+ char buf[100];
- if (trans->iters_live) {
- trace_trans_restart_iters_realloced(trans->ip, trans->size);
- return -EINTR;
+ trans_for_each_iter(trans, iter)
+ printk(KERN_ERR "iter: btree %s pos %s%s%s%s %pS\n",
+ bch2_btree_ids[iter->btree_id],
+ (bch2_bpos_to_text(&PBUF(buf), iter->pos), buf),
+ btree_iter_live(trans, iter) ? " live" : "",
+ (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "",
+ iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "",
+ (void *) iter->ip_allocated);
+
+ trans_for_each_update(trans, i) {
+ char buf[300];
+
+ bch2_bkey_val_to_text(&PBUF(buf), trans->c, bkey_i_to_s_c(i->k));
+ printk(KERN_ERR "update: btree %s %s\n",
+ bch2_btree_ids[i->iter->btree_id], buf);
}
-
- return 0;
+ panic("trans iter oveflow\n");
}
-#endif
static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans)
{
- unsigned idx = __ffs64(~trans->iters_linked);
-
- if (idx < trans->nr_iters)
- goto got_slot;
-
- if (trans->nr_iters == trans->size) {
- struct btree_iter *iter;
-
- BUG_ON(trans->size < BTREE_ITER_MAX);
-
- trans_for_each_iter(trans, iter) {
- pr_err("iter: btree %s pos %llu:%llu%s%s%s %ps",
- bch2_btree_ids[iter->btree_id],
- iter->pos.inode,
- iter->pos.offset,
- (trans->iters_live & (1ULL << iter->idx)) ? " live" : "",
- (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "",
- iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "",
- (void *) iter->ip_allocated);
- }
-
- panic("trans iter oveflow\n");
-#if 0
- ret = bch2_trans_realloc_iters(trans, trans->size * 2);
- if (ret)
- return ERR_PTR(ret);
-#endif
- }
-
- idx = trans->nr_iters++;
- BUG_ON(trans->nr_iters > trans->size);
-
- trans->iters[idx].idx = idx;
-got_slot:
- BUG_ON(trans->iters_linked & (1ULL << idx));
- trans->iters_linked |= 1ULL << idx;
- trans->iters[idx].flags = 0;
- return &trans->iters[idx];
+ struct btree_iter *iter;
+ unsigned idx;
+
+ if (unlikely(trans->iters_linked ==
+ ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
+ btree_trans_iter_alloc_fail(trans);
+
+ idx = __ffs64(~trans->iters_linked);
+ iter = &trans->iters[idx];
+
+ iter->trans = trans;
+ iter->idx = idx;
+ iter->child_idx = U8_MAX;
+ iter->flags = 0;
+ iter->nodes_locked = 0;
+ iter->nodes_intent_locked = 0;
+ trans->iters_linked |= 1ULL << idx;
+ return iter;
}
-static inline void btree_iter_copy(struct btree_iter *dst,
- struct btree_iter *src)
+static void btree_iter_copy(struct btree_iter *dst, struct btree_iter *src)
{
- unsigned i, idx = dst->idx;
+ unsigned i;
- *dst = *src;
- dst->idx = idx;
- dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
+ __bch2_btree_iter_unlock(dst);
+ btree_iter_child_free(dst);
+
+ memcpy(&dst->flags, &src->flags,
+ sizeof(struct btree_iter) - offsetof(struct btree_iter, flags));
for (i = 0; i < BTREE_MAX_DEPTH; i++)
if (btree_node_locked(dst, i))
dst->flags &= ~BTREE_ITER_SET_POS_AFTER_COMMIT;
}
-static inline struct bpos bpos_diff(struct bpos l, struct bpos r)
+struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans,
+ unsigned btree_id, struct bpos pos,
+ unsigned locks_want,
+ unsigned depth,
+ unsigned flags)
{
- if (bkey_cmp(l, r) > 0)
- swap(l, r);
+ struct btree_iter *iter, *best = NULL;
+ struct bpos real_pos, pos_min = POS_MIN;
- return POS(r.inode - l.inode, r.offset - l.offset);
-}
+ if ((flags & BTREE_ITER_TYPE) != BTREE_ITER_NODES &&
+ btree_node_type_is_extents(btree_id) &&
+ !(flags & BTREE_ITER_NOT_EXTENTS) &&
+ !(flags & BTREE_ITER_ALL_SNAPSHOTS))
+ flags |= BTREE_ITER_IS_EXTENTS;
-static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans,
- unsigned btree_id, struct bpos pos,
- unsigned flags)
-{
- struct btree_iter *iter, *best = NULL;
+ if ((flags & BTREE_ITER_TYPE) != BTREE_ITER_NODES &&
+ !btree_type_has_snapshots(btree_id))
+ flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
+
+ if (!(flags & BTREE_ITER_ALL_SNAPSHOTS))
+ pos.snapshot = btree_type_has_snapshots(btree_id)
+ ? U32_MAX : 0;
- BUG_ON(trans->nr_iters > BTREE_ITER_MAX);
+ real_pos = pos;
+
+ if ((flags & BTREE_ITER_IS_EXTENTS) &&
+ bkey_cmp(pos, POS_MAX))
+ real_pos = bpos_nosnap_successor(pos);
trans_for_each_iter(trans, iter) {
if (btree_iter_type(iter) != (flags & BTREE_ITER_TYPE))
if (iter->btree_id != btree_id)
continue;
- if (best &&
- bkey_cmp(bpos_diff(best->pos, pos),
- bpos_diff(iter->pos, pos)) < 0)
- continue;
+ if (best) {
+ int cmp = bkey_cmp(bpos_diff(best->real_pos, real_pos),
+ bpos_diff(iter->real_pos, real_pos));
+
+ if (cmp < 0 ||
+ ((cmp == 0 && btree_iter_keep(trans, iter))))
+ continue;
+ }
best = iter;
}
if (!best) {
iter = btree_trans_iter_alloc(trans);
- if (IS_ERR(iter))
- return iter;
-
- bch2_btree_iter_init(trans, iter, btree_id, pos, flags);
- } else if ((trans->iters_live & (1ULL << best->idx)) ||
- (best->flags & BTREE_ITER_KEEP_UNTIL_COMMIT)) {
+ bch2_btree_iter_init(trans, iter, btree_id);
+ } else if (btree_iter_keep(trans, best)) {
iter = btree_trans_iter_alloc(trans);
- if (IS_ERR(iter))
- return iter;
-
btree_iter_copy(iter, best);
} else {
iter = best;
}
- iter->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
- iter->flags &= ~BTREE_ITER_USER_FLAGS;
- iter->flags |= flags & BTREE_ITER_USER_FLAGS;
+ trans->iters_live |= 1ULL << iter->idx;
+ trans->iters_touched |= 1ULL << iter->idx;
- if (iter->flags & BTREE_ITER_INTENT)
- bch2_btree_iter_upgrade(iter, 1);
- else
- bch2_btree_iter_downgrade(iter);
+ iter->flags = flags;
- BUG_ON(iter->btree_id != btree_id);
- BUG_ON((iter->flags ^ flags) & BTREE_ITER_TYPE);
- BUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
- BUG_ON(iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT);
- BUG_ON(trans->iters_live & (1ULL << iter->idx));
+ iter->snapshot = pos.snapshot;
- trans->iters_live |= 1ULL << iter->idx;
- trans->iters_touched |= 1ULL << iter->idx;
+ /*
+ * If the iterator has locks_want greater than requested, we explicitly
+ * do not downgrade it here - on transaction restart because btree node
+ * split needs to upgrade locks, we might be putting/getting the
+ * iterator again. Downgrading iterators only happens via an explicit
+ * bch2_trans_downgrade().
+ */
- return iter;
-}
+ locks_want = min(locks_want, BTREE_MAX_DEPTH);
+ if (locks_want > iter->locks_want) {
+ iter->locks_want = locks_want;
+ btree_iter_get_locks(iter, true, _THIS_IP_);
+ }
-struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans,
- enum btree_id btree_id,
- struct bpos pos, unsigned flags)
-{
- struct btree_iter *iter =
- __btree_trans_get_iter(trans, btree_id, pos, flags);
+ while (iter->level != depth) {
+ btree_node_unlock(iter, iter->level);
+ iter->l[iter->level].b = BTREE_ITER_NO_NODE_INIT;
+ iter->uptodate = BTREE_ITER_NEED_TRAVERSE;
+ if (iter->level < depth)
+ iter->level++;
+ else
+ iter->level--;
+ }
+
+ iter->min_depth = depth;
+
+ bch2_btree_iter_set_pos(iter, pos);
+ btree_iter_set_search_pos(iter, real_pos);
+
+ trace_trans_get_iter(_RET_IP_, trans->ip,
+ btree_id,
+ &real_pos, locks_want, iter->uptodate,
+ best ? &best->real_pos : &pos_min,
+ best ? best->locks_want : U8_MAX,
+ best ? best->uptodate : U8_MAX);
- if (!IS_ERR(iter))
- __bch2_btree_iter_set_pos(iter, pos,
- btree_node_type_is_extents(btree_id));
return iter;
}
unsigned flags)
{
struct btree_iter *iter =
- __btree_trans_get_iter(trans, btree_id, pos,
- flags|BTREE_ITER_NODES);
- unsigned i;
+ __bch2_trans_get_iter(trans, btree_id, pos,
+ locks_want, depth,
+ BTREE_ITER_NODES|
+ BTREE_ITER_NOT_EXTENTS|
+ BTREE_ITER_ALL_SNAPSHOTS|
+ flags);
- BUG_ON(IS_ERR(iter));
BUG_ON(bkey_cmp(iter->pos, pos));
-
- iter->locks_want = locks_want;
- iter->level = depth;
- iter->min_depth = depth;
-
- for (i = 0; i < ARRAY_SIZE(iter->l); i++)
- iter->l[i].b = NULL;
- iter->l[iter->level].b = BTREE_ITER_NO_NODE_INIT;
+ BUG_ON(iter->locks_want != min(locks_want, BTREE_MAX_DEPTH));
+ BUG_ON(iter->level != depth);
+ BUG_ON(iter->min_depth != depth);
+ iter->ip_allocated = _RET_IP_;
return iter;
}
struct btree_iter *iter;
iter = btree_trans_iter_alloc(trans);
- if (IS_ERR(iter))
- return iter;
-
btree_iter_copy(iter, src);
trans->iters_live |= 1ULL << iter->idx;
* We don't need to preserve this iter since it's cheap to copy it
* again - this will cause trans_iter_put() to free it right away:
*/
- trans->iters_touched &= ~(1ULL << iter->idx);
+ set_btree_iter_dontneed(trans, iter);
return iter;
}
-static int bch2_trans_preload_mem(struct btree_trans *trans, size_t size)
+void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
{
- if (size > trans->mem_bytes) {
+ size_t new_top = trans->mem_top + size;
+ void *p;
+
+ if (new_top > trans->mem_bytes) {
size_t old_bytes = trans->mem_bytes;
- size_t new_bytes = roundup_pow_of_two(size);
- void *new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
+ size_t new_bytes = roundup_pow_of_two(new_top);
+ void *new_mem;
+
+ WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
+
+ new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
+ if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
+ new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
+ new_bytes = BTREE_TRANS_MEM_MAX;
+ kfree(trans->mem);
+ }
if (!new_mem)
- return -ENOMEM;
+ return ERR_PTR(-ENOMEM);
trans->mem = new_mem;
trans->mem_bytes = new_bytes;
if (old_bytes) {
- trace_trans_restart_mem_realloced(trans->ip, new_bytes);
- return -EINTR;
+ trace_trans_restart_mem_realloced(trans->ip, _RET_IP_, new_bytes);
+ return ERR_PTR(-EINTR);
}
}
- return 0;
-}
-
-void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
-{
- void *p;
- int ret;
-
- ret = bch2_trans_preload_mem(trans, trans->mem_top + size);
- if (ret)
- return ERR_PTR(ret);
-
p = trans->mem + trans->mem_top;
trans->mem_top += size;
+ memset(p, 0, size);
return p;
}
{
struct btree_iter *iter;
- trans_for_each_iter(trans, iter)
+ trans_for_each_iter(trans, iter) {
iter->flags &= ~(BTREE_ITER_KEEP_UNTIL_COMMIT|
BTREE_ITER_SET_POS_AFTER_COMMIT);
+ iter->should_be_locked = false;
+ }
bch2_trans_unlink_iters(trans);
trans->iters_touched &= trans->iters_live;
- trans->need_reset = 0;
+ trans->extra_journal_res = 0;
trans->nr_updates = 0;
- trans->nr_updates2 = 0;
trans->mem_top = 0;
+ trans->hooks = NULL;
trans->extra_journal_entries = NULL;
trans->extra_journal_entry_u64s = 0;
(void *) &trans->fs_usage_deltas->memset_start);
}
- if (!(flags & TRANS_RESET_NOTRAVERSE))
+ if (!(flags & TRANS_RESET_NOUNLOCK))
+ bch2_trans_cond_resched(trans);
+
+ if (!(flags & TRANS_RESET_NOTRAVERSE) &&
+ trans->iters_linked)
bch2_btree_iter_traverse_all(trans);
}
static void bch2_trans_alloc_iters(struct btree_trans *trans, struct bch_fs *c)
{
- unsigned new_size = BTREE_ITER_MAX;
- size_t iters_bytes = sizeof(struct btree_iter) * new_size;
- size_t updates_bytes = sizeof(struct btree_insert_entry) * new_size;
+ size_t iters_bytes = sizeof(struct btree_iter) * BTREE_ITER_MAX;
+ size_t updates_bytes = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
void *p = NULL;
BUG_ON(trans->used_mempool);
trans->iters = p; p += iters_bytes;
trans->updates = p; p += updates_bytes;
- trans->updates2 = p; p += updates_bytes;
- trans->size = new_size;
}
void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
unsigned expected_nr_iters,
size_t expected_mem_bytes)
+ __acquires(&c->btree_trans_barrier)
{
memset(trans, 0, sizeof(*trans));
trans->c = c;
if (expected_mem_bytes) {
trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes);
trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL);
+
+ if (!unlikely(trans->mem)) {
+ trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
+ trans->mem_bytes = BTREE_TRANS_MEM_MAX;
+ }
}
trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
}
int bch2_trans_exit(struct btree_trans *trans)
+ __releases(&c->btree_trans_barrier)
{
struct bch_fs *c = trans->c;
bch2_trans_unlock(trans);
#ifdef CONFIG_BCACHEFS_DEBUG
+ if (trans->iters_live) {
+ struct btree_iter *iter;
+
+ trans_for_each_iter(trans, iter)
+ btree_iter_child_free(iter);
+ }
+
+ if (trans->iters_live) {
+ struct btree_iter *iter;
+
+ bch_err(c, "btree iterators leaked!");
+ trans_for_each_iter(trans, iter)
+ if (btree_iter_live(trans, iter))
+ printk(KERN_ERR " btree %s allocated at %pS\n",
+ bch2_btree_ids[iter->btree_id],
+ (void *) iter->ip_allocated);
+ /* Be noisy about this: */
+ bch2_fatal_error(c);
+ }
+
mutex_lock(&trans->c->btree_trans_lock);
list_del(&trans->list);
mutex_unlock(&trans->c->btree_trans_lock);
bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
- kfree(trans->fs_usage_deltas);
- kfree(trans->mem);
+ if (trans->fs_usage_deltas) {
+ if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
+ REPLICAS_DELTA_LIST_MAX)
+ mempool_free(trans->fs_usage_deltas,
+ &trans->c->replicas_delta_pool);
+ else
+ kfree(trans->fs_usage_deltas);
+ }
+
+ if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
+ mempool_free(trans->mem, &trans->c->btree_trans_mem_pool);
+ else
+ kfree(trans->mem);
#ifdef __KERNEL__
/*
*/
trans->iters = this_cpu_xchg(c->btree_iters_bufs->iter, trans->iters);
#endif
+
if (trans->iters)
mempool_free(trans->iters, &trans->c->btree_iters_pool);
struct btree_bkey_cached_common *_b,
enum btree_iter_type type)
{
- pr_buf(out, " %px l=%u %s:",
- _b, _b->level, bch2_btree_ids[_b->btree_id]);
+ pr_buf(out, " l=%u %s:",
+ _b->level, bch2_btree_ids[_b->btree_id]);
bch2_bpos_to_text(out, btree_node_pos(_b, type));
}
+#ifdef CONFIG_BCACHEFS_DEBUG
+static bool trans_has_btree_nodes_locked(struct btree_trans *trans)
+{
+ struct btree_iter *iter;
+
+ trans_for_each_iter(trans, iter)
+ if (btree_iter_type(iter) != BTREE_ITER_CACHED &&
+ iter->nodes_locked)
+ return true;
+ return false;
+}
+#endif
+
void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
{
#ifdef CONFIG_BCACHEFS_DEBUG
mutex_lock(&c->btree_trans_lock);
list_for_each_entry(trans, &c->btree_trans_list, list) {
- pr_buf(out, "%i %px %ps\n", trans->pid, trans, (void *) trans->ip);
+ if (!trans_has_btree_nodes_locked(trans))
+ continue;
+
+ pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip);
trans_for_each_iter(trans, iter) {
if (!iter->nodes_locked)
continue;
- pr_buf(out, " iter %u %s:",
+ pr_buf(out, " iter %u %c %s:",
iter->idx,
+ btree_iter_type(iter) == BTREE_ITER_CACHED ? 'c' : 'b',
bch2_btree_ids[iter->btree_id]);
bch2_bpos_to_text(out, iter->pos);
pr_buf(out, "\n");
b = READ_ONCE(trans->locking);
if (b) {
- pr_buf(out, " locking iter %u l=%u %s:",
+ iter = &trans->iters[trans->locking_iter_idx];
+ pr_buf(out, " locking iter %u %c l=%u %s:",
trans->locking_iter_idx,
+ btree_iter_type(iter) == BTREE_ITER_CACHED ? 'c' : 'b',
trans->locking_level,
bch2_btree_ids[trans->locking_btree_id]);
bch2_bpos_to_text(out, trans->locking_pos);
-
pr_buf(out, " node ");
bch2_btree_iter_node_to_text(out,
(void *) b,
- btree_iter_type(&trans->iters[trans->locking_iter_idx]));
+ btree_iter_type(iter));
pr_buf(out, "\n");
}
}
void bch2_fs_btree_iter_exit(struct bch_fs *c)
{
+ mempool_exit(&c->btree_trans_mem_pool);
mempool_exit(&c->btree_iters_pool);
cleanup_srcu_struct(&c->btree_trans_barrier);
}
return init_srcu_struct(&c->btree_trans_barrier) ?:
mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
sizeof(struct btree_iter) * nr +
- sizeof(struct btree_insert_entry) * nr +
- sizeof(struct btree_insert_entry) * nr);
+ sizeof(struct btree_insert_entry) * nr) ?:
+ mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
+ BTREE_TRANS_MEM_MAX);
}
projects / bcachefs-tools-debian / blobdiff