X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_update_leaf.c;h=ac6a039b3fa45f867c5a7eeb04f4516611bc13f6;hb=2e2d5a3f7ed4f6d4e2a45011a82386ba82cd3976;hp=007aa5ef40910e89c6b764500e5d9e3055d0631b;hpb=807abf36c1e119825d42cda6f6b249649ca44eb5;p=bcachefs-tools-debian

diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_update_leaf.c
index 007aa5e..ac6a039 100644
--- a/libbcachefs/btree_update_leaf.c
+++ b/libbcachefs/btree_update_leaf.c
@@ -1,18 +1,61 @@
+// SPDX-License-Identifier: GPL-2.0
 
 #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_key_cache.h"
 #include "btree_locking.h"
+#include "buckets.h"
 #include "debug.h"
-#include "extents.h"
+#include "error.h"
+#include "extent_update.h"
 #include "journal.h"
+#include "journal_reclaim.h"
 #include "keylist.h"
+#include "replicas.h"
 
+#include <linux/prefetch.h>
 #include <linux/sort.h>
 #include <trace/events/bcachefs.h>
 
+static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
+					 const struct btree_insert_entry *r)
+{
+	return   cmp_int(l->btree_id,	r->btree_id) ?:
+		 -cmp_int(l->level,	r->level) ?:
+		 bpos_cmp(l->k->k.p,	r->k->k.p);
+}
+
+static inline bool same_leaf_as_prev(struct btree_trans *trans,
+				     struct btree_insert_entry *i)
+{
+	return i != trans->updates2 &&
+		iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
+}
+
+inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
+					    struct btree_iter *iter)
+{
+	bch2_btree_node_lock_write(b, iter);
+
+	if (btree_iter_type(iter) == BTREE_ITER_CACHED)
+		return;
+
+	if (unlikely(btree_node_just_written(b)) &&
+	    bch2_btree_post_write_cleanup(c, b))
+		bch2_btree_iter_reinit_node(iter, b);
+
+	/*
+	 * 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);
+}
+
 /* Inserting into a given leaf node (last stage of insert): */
 
 /* Handle overwrites and do insert, for non extents: */
@@ -21,193 +64,143 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 				struct btree_node_iter *node_iter,
 				struct bkey_i *insert)
 {
-	const struct bkey_format *f = &b->format;
 	struct bkey_packed *k;
-	struct bset_tree *t;
-	unsigned clobber_u64s;
+	unsigned clobber_u64s = 0, new_u64s = 0;
 
 	EBUG_ON(btree_node_just_written(b));
 	EBUG_ON(bset_written(b, btree_bset_last(b)));
 	EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
-	EBUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0 ||
-		bkey_cmp(insert->k.p, b->data->max_key) > 0);
+	EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
+	EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
+	EBUG_ON(insert->k.u64s >
+		bch_btree_keys_u64s_remaining(iter->trans->c, b));
+	EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
 
 	k = bch2_btree_node_iter_peek_all(node_iter, b);
-	if (k && !bkey_cmp_packed(b, k, &insert->k)) {
-		BUG_ON(bkey_whiteout(k));
-
-		t = bch2_bkey_to_bset(b, k);
-
-		if (bset_unwritten(b, bset(b, t)) &&
-		    bkey_val_u64s(&insert->k) == bkeyp_val_u64s(f, k) &&
-		    !bkey_whiteout(&insert->k)) {
-			k->type = insert->k.type;
-			memcpy_u64s(bkeyp_val(f, k), &insert->v,
-				    bkey_val_u64s(&insert->k));
-			return true;
-		}
+	if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
+		k = NULL;
 
-		insert->k.needs_whiteout = k->needs_whiteout;
+	/* @k is the key being overwritten/deleted, if any: */
+	EBUG_ON(k && bkey_deleted(k));
 
-		btree_keys_account_key_drop(&b->nr, t - b->set, k);
+	/* Deleting, but not found? nothing to do: */
+	if (bkey_deleted(&insert->k) && !k)
+		return false;
 
-		if (t == bset_tree_last(b)) {
-			clobber_u64s = k->u64s;
+	if (bkey_deleted(&insert->k)) {
+		/* Deleting: */
+		btree_account_key_drop(b, k);
+		k->type = KEY_TYPE_deleted;
 
-			/*
-			 * If we're deleting, and the key we're deleting doesn't
-			 * need a whiteout (it wasn't overwriting a key that had
-			 * been written to disk) - just delete it:
-			 */
-			if (bkey_whiteout(&insert->k) && !k->needs_whiteout) {
-				bch2_bset_delete(b, k, clobber_u64s);
-				bch2_btree_node_iter_fix(iter, b, node_iter, t,
-							k, clobber_u64s, 0);
-				return true;
-			}
+		if (k->needs_whiteout)
+			push_whiteout(iter->trans->c, b, insert->k.p);
+		k->needs_whiteout = false;
 
-			goto overwrite;
+		if (k >= btree_bset_last(b)->start) {
+			clobber_u64s = k->u64s;
+			bch2_bset_delete(b, k, clobber_u64s);
+			goto fix_iter;
+		} else {
+			bch2_btree_iter_fix_key_modified(iter, b, k);
 		}
 
-		k->type = KEY_TYPE_DELETED;
-		bch2_btree_node_iter_fix(iter, b, node_iter, t, k,
-					k->u64s, k->u64s);
+		return true;
+	}
+
+	if (k) {
+		/* Overwriting: */
+		btree_account_key_drop(b, k);
+		k->type = KEY_TYPE_deleted;
 
-		if (bkey_whiteout(&insert->k)) {
-			reserve_whiteout(b, t, k);
-			return true;
+		insert->k.needs_whiteout = k->needs_whiteout;
+		k->needs_whiteout = false;
+
+		if (k >= btree_bset_last(b)->start) {
+			clobber_u64s = k->u64s;
+			goto overwrite;
 		} else {
-			k->needs_whiteout = false;
+			bch2_btree_iter_fix_key_modified(iter, b, k);
 		}
-	} else {
-		/*
-		 * Deleting, but the key to delete wasn't found - nothing to do:
-		 */
-		if (bkey_whiteout(&insert->k))
-			return false;
-
-		insert->k.needs_whiteout = false;
 	}
 
-	t = bset_tree_last(b);
-	k = bch2_btree_node_iter_bset_pos(node_iter, b, t);
-	clobber_u64s = 0;
+	k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 overwrite:
 	bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
-	if (k->u64s != clobber_u64s || bkey_whiteout(&insert->k))
-		bch2_btree_node_iter_fix(iter, b, node_iter, t, k,
-					clobber_u64s, k->u64s);
+	new_u64s = k->u64s;
+fix_iter:
+	if (clobber_u64s != new_u64s)
+		bch2_btree_node_iter_fix(iter, b, node_iter, k,
+					 clobber_u64s, new_u64s);
 	return true;
 }
 
-static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
+static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 			       unsigned i, u64 seq)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	struct btree_write *w = container_of(pin, struct btree_write, journal);
 	struct btree *b = container_of(w, struct btree, writes[i]);
 
-	six_lock_read(&b->lock);
+	btree_node_lock_type(c, b, SIX_LOCK_read);
 	bch2_btree_node_write_cond(c, b,
-			(btree_current_write(b) == w &&
-			 w->journal.pin_list == journal_seq_pin(j, seq)));
-	six_unlock_read(&b->lock);
+		(btree_current_write(b) == w && w->journal.seq == seq));
+	six_unlock_read(&b->c.lock);
+	return 0;
 }
 
-static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
+static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 {
 	return __btree_node_flush(j, pin, 0, seq);
 }
 
-static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
+static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 {
 	return __btree_node_flush(j, pin, 1, seq);
 }
 
-void bch2_btree_journal_key(struct btree_insert *trans,
-			   struct btree_iter *iter,
-			   struct bkey_i *insert)
+inline void bch2_btree_add_journal_pin(struct bch_fs *c,
+				       struct btree *b, u64 seq)
 {
-	struct bch_fs *c = trans->c;
-	struct journal *j = &c->journal;
-	struct btree *b = iter->l[0].b;
 	struct btree_write *w = btree_current_write(b);
 
-	EBUG_ON(iter->level || b->level);
-	EBUG_ON(trans->journal_res.ref !=
-		!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
-
-	if (likely(trans->journal_res.ref)) {
-		u64 seq = trans->journal_res.seq;
-		bool needs_whiteout = insert->k.needs_whiteout;
-
-		/* ick */
-		insert->k.needs_whiteout = false;
-		bch2_journal_add_keys(j, &trans->journal_res,
-				      iter->btree_id, insert);
-		insert->k.needs_whiteout = needs_whiteout;
-
-		bch2_journal_set_has_inode(j, &trans->journal_res,
-					   insert->k.p.inode);
-
-		if (trans->journal_seq)
-			*trans->journal_seq = seq;
-		btree_bset_last(b)->journal_seq = cpu_to_le64(seq);
-	}
-
-	if (unlikely(!journal_pin_active(&w->journal)))
-		bch2_journal_pin_add(j, &trans->journal_res,
-				     &w->journal,
-				     btree_node_write_idx(b) == 0
-				     ? btree_node_flush0
-				     : btree_node_flush1);
-
-	if (unlikely(!btree_node_dirty(b)))
-		set_btree_node_dirty(b);
-}
-
-static enum btree_insert_ret
-bch2_insert_fixup_key(struct btree_insert *trans,
-		     struct btree_insert_entry *insert)
-{
-	struct btree_iter *iter = insert->iter;
-	struct btree_iter_level *l = &iter->l[0];
-
-	EBUG_ON(iter->level);
-	EBUG_ON(insert->k->k.u64s >
-		bch_btree_keys_u64s_remaining(trans->c, l->b));
-
-	if (bch2_btree_bset_insert_key(iter, l->b, &l->iter,
-				       insert->k))
-		bch2_btree_journal_key(trans, iter, insert->k);
-
-	trans->did_work = true;
-	return BTREE_INSERT_OK;
+	bch2_journal_pin_add(&c->journal, seq, &w->journal,
+			     btree_node_write_idx(b) == 0
+			     ? btree_node_flush0
+			     : btree_node_flush1);
 }
 
 /**
  * btree_insert_key - insert a key one key into a leaf node
  */
-static enum btree_insert_ret
-btree_insert_key_leaf(struct btree_insert *trans,
-		      struct btree_insert_entry *insert)
+static bool btree_insert_key_leaf(struct btree_trans *trans,
+				  struct btree_iter *iter,
+				  struct bkey_i *insert)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter *iter = insert->iter;
-	struct btree *b = iter->l[0].b;
-	enum btree_insert_ret ret;
-	int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
+	struct btree *b = iter_l(iter)->b;
+	struct bset_tree *t = bset_tree_last(b);
+	struct bset *i = bset(b, t);
+	int old_u64s = bset_u64s(t);
 	int old_live_u64s = b->nr.live_u64s;
 	int live_u64s_added, u64s_added;
 
-	btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
+	EBUG_ON(!iter->level &&
+		!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 
-	ret = !btree_node_is_extents(b)
-		? bch2_insert_fixup_key(trans, insert)
-		: bch2_insert_fixup_extent(trans, insert);
+	if (unlikely(!bch2_btree_bset_insert_key(iter, b,
+					&iter_l(iter)->iter, insert)))
+		return false;
+
+	i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
+					 le64_to_cpu(i->journal_seq)));
+
+	bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
+
+	if (unlikely(!btree_node_dirty(b)))
+		set_btree_node_dirty(c, b);
 
 	live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
-	u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
+	u64s_added = (int) bset_u64s(t) - old_u64s;
 
 	if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 		b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
@@ -218,357 +211,997 @@ btree_insert_key_leaf(struct btree_insert *trans,
 	    bch2_maybe_compact_whiteouts(c, b))
 		bch2_btree_iter_reinit_node(iter, b);
 
-	trace_btree_insert_key(c, b, insert->k);
-	return ret;
+	trace_btree_insert_key(c, b, insert);
+	return true;
 }
 
-static bool same_leaf_as_prev(struct btree_insert *trans,
-			      struct btree_insert_entry *i)
+/* Cached btree updates: */
+
+/* Normal update interface: */
+
+static inline void btree_insert_entry_checks(struct btree_trans *trans,
+					     struct btree_insert_entry *i)
 {
-	/*
-	 * Because we sorted the transaction entries, if multiple iterators
-	 * point to the same leaf node they'll always be adjacent now:
-	 */
-	return i != trans->entries &&
-		i[0].iter->l[0].b == i[-1].iter->l[0].b;
+	BUG_ON(!i->is_extent && bpos_cmp(i->k->k.p, i->iter->real_pos));
+	BUG_ON(i->level		!= i->iter->level);
+	BUG_ON(i->btree_id	!= i->iter->btree_id);
 }
 
-#define trans_for_each_entry(trans, i)					\
-	for ((i) = (trans)->entries; (i) < (trans)->entries + (trans)->nr; (i)++)
+static noinline int
+bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
+{
+	struct bch_fs *c = trans->c;
+	int ret;
 
-inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
-					    struct btree_iter *iter)
+	bch2_trans_unlock(trans);
+
+	ret = bch2_journal_preres_get(&c->journal,
+			&trans->journal_preres, u64s, 0);
+	if (ret)
+		return ret;
+
+	if (!bch2_trans_relock(trans)) {
+		trace_trans_restart_journal_preres_get(trans->ip);
+		return -EINTR;
+	}
+
+	return 0;
+}
+
+static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
+					     unsigned flags)
 {
-	bch2_btree_node_lock_write(b, iter);
+	struct bch_fs *c = trans->c;
+	int ret;
 
-	if (btree_node_just_written(b) &&
-	    bch2_btree_post_write_cleanup(c, b))
-		bch2_btree_iter_reinit_node(iter, b);
+	if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
+		flags |= JOURNAL_RES_GET_RESERVED;
+
+	ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
+				   trans->journal_u64s, flags);
+
+	return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
+}
+
+static enum btree_insert_ret
+btree_key_can_insert(struct btree_trans *trans,
+		     struct btree_iter *iter,
+		     unsigned u64s)
+{
+	struct bch_fs *c = trans->c;
+	struct btree *b = iter_l(iter)->b;
+
+	if (!bch2_btree_node_insert_fits(c, b, u64s))
+		return BTREE_INSERT_BTREE_NODE_FULL;
+
+	return BTREE_INSERT_OK;
+}
+
+static enum btree_insert_ret
+btree_key_can_insert_cached(struct btree_trans *trans,
+			    struct btree_iter *iter,
+			    unsigned u64s)
+{
+	struct bkey_cached *ck = (void *) iter->l[0].b;
+	unsigned new_u64s;
+	struct bkey_i *new_k;
+
+	BUG_ON(iter->level);
+
+	if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
+	    bch2_btree_key_cache_must_wait(trans->c) &&
+	    !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
+		return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 
 	/*
-	 * If the last bset has been written, or if it's gotten too big - start
-	 * a new bset to insert into:
+	 * bch2_varint_decode can read past the end of the buffer by at most 7
+	 * bytes (it won't be used):
 	 */
-	if (want_new_bset(c, b))
-		bch2_btree_init_next(c, b, iter);
+	u64s += 1;
+
+	if (u64s <= ck->u64s)
+		return BTREE_INSERT_OK;
+
+	new_u64s	= roundup_pow_of_two(u64s);
+	new_k		= krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
+	if (!new_k)
+		return -ENOMEM;
+
+	ck->u64s	= new_u64s;
+	ck->k		= new_k;
+	return BTREE_INSERT_OK;
+}
+
+static inline void do_btree_insert_one(struct btree_trans *trans,
+				       struct btree_insert_entry *i)
+{
+	struct bch_fs *c = trans->c;
+	struct journal *j = &c->journal;
+	bool did_work;
+
+	EBUG_ON(trans->journal_res.ref !=
+		!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
+
+	i->k->k.needs_whiteout = false;
+
+	did_work = (btree_iter_type(i->iter) != BTREE_ITER_CACHED)
+		? btree_insert_key_leaf(trans, i->iter, i->k)
+		: bch2_btree_insert_key_cached(trans, i->iter, i->k);
+	if (!did_work)
+		return;
+
+	if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
+		bch2_journal_add_keys(j, &trans->journal_res,
+				      i->btree_id,
+				      i->level,
+				      i->k);
+
+		bch2_journal_set_has_inode(j, &trans->journal_res,
+					   i->k->k.p.inode);
+
+		if (trans->journal_seq)
+			*trans->journal_seq = trans->journal_res.seq;
+	}
 }
 
-static void multi_lock_write(struct bch_fs *c, struct btree_insert *trans)
+static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 {
+	__bch2_btree_iter_unlock(iter);
+}
+
+static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
+{
+	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
 
-	trans_for_each_entry(trans, i)
-		if (!same_leaf_as_prev(trans, i))
-			bch2_btree_node_lock_for_insert(c, i->iter->l[0].b,
-							i->iter);
+	trans_for_each_update(trans, i) {
+		/*
+		 * XXX: synchronization of cached update triggers with gc
+		 */
+		BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
+
+		if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
+			bch2_mark_update(trans, i->iter, i->k, NULL,
+					 i->trigger_flags|BTREE_TRIGGER_GC);
+	}
 }
 
-static void multi_unlock_write(struct btree_insert *trans)
+static inline int
+bch2_trans_commit_write_locked(struct btree_trans *trans,
+			       struct btree_insert_entry **stopped_at)
 {
+	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
+	struct btree_trans_commit_hook *h;
+	unsigned u64s = 0;
+	bool marking = false;
+	int ret;
 
-	trans_for_each_entry(trans, i)
+	if (race_fault()) {
+		trace_trans_restart_fault_inject(trans->ip);
+		return -EINTR;
+	}
+
+	/*
+	 * Check if the insert will fit in the leaf node with the write lock
+	 * held, otherwise another thread could write the node changing the
+	 * amount of space available:
+	 */
+
+	prefetch(&trans->c->journal.flags);
+
+	h = trans->hooks;
+	while (h) {
+		ret = h->fn(trans, h);
+		if (ret)
+			return ret;
+		h = h->next;
+	}
+
+	trans_for_each_update2(trans, i) {
+		/* Multiple inserts might go to same leaf: */
 		if (!same_leaf_as_prev(trans, i))
-			bch2_btree_node_unlock_write(i->iter->l[0].b, i->iter);
+			u64s = 0;
+
+		u64s += i->k->k.u64s;
+		ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
+			? btree_key_can_insert(trans, i->iter, u64s)
+			: btree_key_can_insert_cached(trans, i->iter, u64s);
+		if (ret) {
+			*stopped_at = i;
+			return ret;
+		}
+
+		if (btree_node_type_needs_gc(i->bkey_type))
+			marking = true;
+	}
+
+	if (marking) {
+		percpu_down_read(&c->mark_lock);
+	}
+
+	/* Must be called under mark_lock: */
+	if (marking && trans->fs_usage_deltas &&
+	    !bch2_replicas_delta_list_marked(c, trans->fs_usage_deltas)) {
+		ret = BTREE_INSERT_NEED_MARK_REPLICAS;
+		goto err;
+	}
+
+	/*
+	 * Don't get journal reservation until after we know insert will
+	 * succeed:
+	 */
+	if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
+		ret = bch2_trans_journal_res_get(trans,
+				JOURNAL_RES_GET_NONBLOCK);
+		if (ret)
+			goto err;
+	} else {
+		trans->journal_res.seq = c->journal.replay_journal_seq;
+	}
+
+	if (unlikely(trans->extra_journal_entry_u64s)) {
+		memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
+				  trans->extra_journal_entries,
+				  trans->extra_journal_entry_u64s);
+
+		trans->journal_res.offset	+= trans->extra_journal_entry_u64s;
+		trans->journal_res.u64s		-= trans->extra_journal_entry_u64s;
+	}
+
+	/*
+	 * Not allowed to fail after we've gotten our journal reservation - we
+	 * have to use it:
+	 */
+
+	if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
+		if (bch2_journal_seq_verify)
+			trans_for_each_update2(trans, i)
+				i->k->k.version.lo = trans->journal_res.seq;
+		else if (bch2_inject_invalid_keys)
+			trans_for_each_update2(trans, i)
+				i->k->k.version = MAX_VERSION;
+	}
+
+	trans_for_each_update(trans, i)
+		if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
+			bch2_mark_update(trans, i->iter, i->k,
+					 NULL, i->trigger_flags);
+
+	if (marking && trans->fs_usage_deltas)
+		bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas);
+
+	if (unlikely(c->gc_pos.phase))
+		bch2_trans_mark_gc(trans);
+
+	trans_for_each_update2(trans, i)
+		do_btree_insert_one(trans, i);
+err:
+	if (marking) {
+		percpu_up_read(&c->mark_lock);
+	}
+
+	return ret;
 }
 
-static inline int btree_trans_cmp(struct btree_insert_entry l,
-				  struct btree_insert_entry r)
+static noinline int maybe_do_btree_merge(struct btree_trans *trans, struct btree_iter *iter)
 {
-	return btree_iter_cmp(l.iter, r.iter);
-}
+	struct btree_insert_entry *i;
+	struct btree *b = iter_l(iter)->b;
+	struct bkey_s_c old;
+	int u64s_delta = 0;
+	int ret;
 
-/* Normal update interface: */
+	/*
+	 * Inserting directly into interior nodes is an uncommon operation with
+	 * various weird edge cases: also, a lot of things about
+	 * BTREE_ITER_NODES iters need to be audited
+	 */
+	if (unlikely(btree_iter_type(iter) != BTREE_ITER_KEYS))
+		return 0;
 
-/**
- * __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
+	BUG_ON(iter->level);
+
+	trans_for_each_update2(trans, i) {
+		if (iter_l(i->iter)->b != b)
+			continue;
+
+		old = bch2_btree_iter_peek_slot(i->iter);
+		ret = bkey_err(old);
+		if (ret)
+			return ret;
+
+		u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
+		u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
+	}
+
+	return u64s_delta <= 0
+		? (bch2_foreground_maybe_merge(trans->c, iter, iter->level,
+				trans->flags & ~BTREE_INSERT_NOUNLOCK) ?: -EINTR)
+		: 0;
+}
+
+/*
+ * Get journal reservation, take write locks, and attempt to do btree update(s):
  */
-int __bch2_btree_insert_at(struct btree_insert *trans)
+static inline int do_bch2_trans_commit(struct btree_trans *trans,
+				       struct btree_insert_entry **stopped_at)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
-	struct btree_iter *split = NULL;
-	bool cycle_gc_lock = false;
-	unsigned u64s;
+	struct btree_iter *iter;
 	int ret;
 
-	trans_for_each_entry(trans, i) {
-		BUG_ON(i->iter->level);
-		BUG_ON(bkey_cmp(bkey_start_pos(&i->k->k), i->iter->pos));
-		BUG_ON(debug_check_bkeys(c) &&
-		       bch2_bkey_invalid(c, i->iter->btree_id,
-					 bkey_i_to_s_c(i->k)));
-	}
+	trans_for_each_update2(trans, i) {
+		struct btree *b;
 
-	bubble_sort(trans->entries, trans->nr, btree_trans_cmp);
+		BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 
-	if (unlikely(!percpu_ref_tryget(&c->writes)))
-		return -EROFS;
-retry_locks:
-	ret = -EINTR;
-	trans_for_each_entry(trans, i) {
-		if (!bch2_btree_iter_set_locks_want(i->iter, 1))
-			goto err;
+		if (btree_iter_type(i->iter) == BTREE_ITER_CACHED)
+			continue;
 
-		if (i->iter->uptodate == BTREE_ITER_NEED_TRAVERSE) {
-			ret = bch2_btree_iter_traverse(i->iter);
-			if (ret)
-				goto err;
+		b = iter_l(i->iter)->b;
+		if (b->sib_u64s[0] < c->btree_foreground_merge_threshold ||
+		    b->sib_u64s[1] < c->btree_foreground_merge_threshold) {
+			ret = maybe_do_btree_merge(trans, i->iter);
+			if (unlikely(ret))
+				return ret;
 		}
 	}
-retry:
-	trans->did_work = false;
-	u64s = 0;
-	trans_for_each_entry(trans, i)
-		if (!i->done)
-			u64s += jset_u64s(i->k->k.u64s + i->extra_res);
 
-	memset(&trans->journal_res, 0, sizeof(trans->journal_res));
+	trans_for_each_update2(trans, i)
+		BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 
-	ret = !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)
-		? bch2_journal_res_get(&c->journal,
-				      &trans->journal_res,
-				      u64s, u64s)
-		: 0;
-	if (ret)
-		goto err;
+	ret = bch2_journal_preres_get(&c->journal,
+			&trans->journal_preres, trans->journal_preres_u64s,
+			JOURNAL_RES_GET_NONBLOCK|
+			((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
+			 ? JOURNAL_RES_GET_RESERVED : 0));
+	if (unlikely(ret == -EAGAIN))
+		ret = bch2_trans_journal_preres_get_cold(trans,
+						trans->journal_preres_u64s);
+	if (unlikely(ret))
+		return ret;
+
+	/*
+	 * Can't be holding any read locks when we go to take write locks:
+	 * another thread could be holding an intent lock on the same node we
+	 * have a read lock on, and it'll block trying to take a write lock
+	 * (because we hold a read lock) and it could be blocking us by holding
+	 * its own read lock (while we're trying to to take write locks).
+	 *
+	 * note - this must be done after bch2_trans_journal_preres_get_cold()
+	 * or anything else that might call bch2_trans_relock(), since that
+	 * would just retake the read locks:
+	 */
+	trans_for_each_iter(trans, iter) {
+		if (iter->nodes_locked != iter->nodes_intent_locked) {
+			if (btree_iter_keep(trans, iter)) {
+				if (!bch2_btree_iter_upgrade(iter, 1)) {
+					trace_trans_restart_upgrade(trans->ip);
+					return -EINTR;
+				}
+			} else {
+				bch2_btree_iter_unlock_noinline(iter);
+			}
+		}
+	}
 
-	multi_lock_write(c, trans);
+	trans_for_each_update2(trans, i) {
+		const char *invalid = bch2_bkey_invalid(c,
+				bkey_i_to_s_c(i->k), i->bkey_type);
+		if (invalid) {
+			char buf[200];
 
-	if (race_fault()) {
-		ret = -EINTR;
-		goto unlock;
+			bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
+			bch_err(c, "invalid bkey %s on insert: %s\n", buf, invalid);
+			bch2_fatal_error(c);
+		}
+		btree_insert_entry_checks(trans, i);
 	}
+	bch2_btree_trans_verify_locks(trans);
 
-	u64s = 0;
-	trans_for_each_entry(trans, i) {
-		/* Multiple inserts might go to same leaf: */
+	trans_for_each_update2(trans, i)
 		if (!same_leaf_as_prev(trans, i))
-			u64s = 0;
+			bch2_btree_node_lock_for_insert(c,
+					iter_l(i->iter)->b, i->iter);
+
+	ret = bch2_trans_commit_write_locked(trans, stopped_at);
+
+	trans_for_each_update2(trans, i)
+		if (!same_leaf_as_prev(trans, i))
+			bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
+							     i->iter);
+
+	if (!ret && trans->journal_pin)
+		bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
+				     trans->journal_pin, NULL);
+
+	/*
+	 * Drop journal reservation after dropping write locks, since dropping
+	 * the journal reservation may kick off a journal write:
+	 */
+	bch2_journal_res_put(&c->journal, &trans->journal_res);
+
+	if (unlikely(ret))
+		return ret;
+
+	bch2_trans_downgrade(trans);
+
+	return 0;
+}
+
+static int journal_reclaim_wait_done(struct bch_fs *c)
+{
+	int ret = bch2_journal_error(&c->journal) ?:
+		!bch2_btree_key_cache_must_wait(c);
+
+	if (!ret)
+		journal_reclaim_kick(&c->journal);
+	return ret;
+}
+
+static noinline
+int bch2_trans_commit_error(struct btree_trans *trans,
+			    struct btree_insert_entry *i,
+			    int ret)
+{
+	struct bch_fs *c = trans->c;
+	unsigned flags = trans->flags;
+
+	/*
+	 * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
+	 * update; if we haven't done anything yet it doesn't apply
+	 */
+	flags &= ~BTREE_INSERT_NOUNLOCK;
+
+	switch (ret) {
+	case BTREE_INSERT_BTREE_NODE_FULL:
+		ret = bch2_btree_split_leaf(c, i->iter, flags);
 
 		/*
-		 * bch2_btree_node_insert_fits() must be called under write lock:
-		 * with only an intent lock, another thread can still call
-		 * bch2_btree_node_write(), converting an unwritten bset to a
-		 * written one
+		 * if the split succeeded without dropping locks the insert will
+		 * still be atomic (what the caller peeked() and is overwriting
+		 * won't have changed)
 		 */
-		if (!i->done) {
-			u64s += i->k->k.u64s + i->extra_res;
-			if (!bch2_btree_node_insert_fits(c,
-					i->iter->l[0].b, u64s)) {
-				split = i->iter;
-				goto unlock;
-			}
+#if 0
+		/*
+		 * 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)
+			goto retry;
+#endif
+
+		/*
+		 * don't care if we got ENOSPC because we told split it
+		 * couldn't block:
+		 */
+		if (!ret ||
+		    ret == -EINTR ||
+		    (flags & BTREE_INSERT_NOUNLOCK)) {
+			trace_trans_restart_btree_node_split(trans->ip);
+			ret = -EINTR;
 		}
+		break;
+	case BTREE_INSERT_ENOSPC:
+		ret = -ENOSPC;
+		break;
+	case BTREE_INSERT_NEED_MARK_REPLICAS:
+		bch2_trans_unlock(trans);
+
+		ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
+		if (ret)
+			return ret;
+
+		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);
+
+		if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
+		    !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED))
+			return -EAGAIN;
+
+		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;
+	case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
+		bch2_trans_unlock(trans);
+
+		trace_trans_blocked_journal_reclaim(trans->ip);
+
+		wait_event_freezable(c->journal.reclaim_wait,
+				     (ret = journal_reclaim_wait_done(c)));
+		if (ret < 0)
+			return ret;
+
+		if (bch2_trans_relock(trans))
+			return 0;
+
+		trace_trans_restart_journal_reclaim(trans->ip);
+		ret = -EINTR;
+		break;
+	default:
+		BUG_ON(ret >= 0);
+		break;
 	}
 
-	ret = 0;
-	split = NULL;
-	cycle_gc_lock = false;
+	return ret;
+}
 
-	trans_for_each_entry(trans, i) {
-		if (i->done)
-			continue;
+static noinline int
+bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
+{
+	struct bch_fs *c = trans->c;
+	int ret;
 
-		switch (btree_insert_key_leaf(trans, i)) {
-		case BTREE_INSERT_OK:
-			i->done = true;
-			break;
-		case BTREE_INSERT_JOURNAL_RES_FULL:
-		case BTREE_INSERT_NEED_TRAVERSE:
-			ret = -EINTR;
-			break;
-		case BTREE_INSERT_NEED_RESCHED:
-			ret = -EAGAIN;
-			break;
-		case BTREE_INSERT_BTREE_NODE_FULL:
-			split = i->iter;
-			break;
-		case BTREE_INSERT_ENOSPC:
-			ret = -ENOSPC;
+	if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
+		return -EROFS;
+
+	bch2_trans_unlock(trans);
+
+	ret = bch2_fs_read_write_early(c);
+	if (ret)
+		return ret;
+
+	percpu_ref_get(&c->writes);
+	return 0;
+}
+
+static void __bch2_trans_update2(struct btree_trans *trans,
+				 struct btree_insert_entry n)
+{
+	struct btree_insert_entry *i;
+
+	btree_insert_entry_checks(trans, &n);
+
+	EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX);
+
+	n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+
+	trans_for_each_update2(trans, i)
+		if (btree_insert_entry_cmp(&n, i) <= 0)
 			break;
-		case BTREE_INSERT_NEED_GC_LOCK:
-			cycle_gc_lock = true;
-			ret = -EINTR;
+
+	if (i < trans->updates2 + trans->nr_updates2 &&
+	    !btree_insert_entry_cmp(&n, i))
+		*i = n;
+	else
+		array_insert_item(trans->updates2, trans->nr_updates2,
+				  i - trans->updates2, n);
+}
+
+static void bch2_trans_update2(struct btree_trans *trans,
+			       struct btree_iter *iter,
+			       struct bkey_i *insert)
+{
+	__bch2_trans_update2(trans, (struct btree_insert_entry) {
+		.bkey_type	= __btree_node_type(iter->level, iter->btree_id),
+		.btree_id	= iter->btree_id,
+		.level		= iter->level,
+		.iter		= iter,
+		.k		= insert,
+	});
+}
+
+static int extent_update_to_keys(struct btree_trans *trans,
+				 struct btree_insert_entry n)
+{
+	int ret;
+
+	ret = bch2_extent_can_insert(trans, n.iter, n.k);
+	if (ret)
+		return ret;
+
+	if (bkey_deleted(&n.k->k))
+		return 0;
+
+	n.iter = bch2_trans_get_iter(trans, n.iter->btree_id, n.k->k.p,
+				     BTREE_ITER_INTENT|
+				     BTREE_ITER_NOT_EXTENTS);
+	n.is_extent = false;
+
+	__bch2_trans_update2(trans, n);
+	bch2_trans_iter_put(trans, n.iter);
+	return 0;
+}
+
+static int extent_handle_overwrites(struct btree_trans *trans,
+				    enum btree_id btree_id,
+				    struct bkey_i *insert)
+{
+	struct btree_iter *iter, *update_iter;
+	struct bpos start = bkey_start_pos(&insert->k);
+	struct bkey_i *update;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	iter = bch2_trans_get_iter(trans, btree_id, start,
+				   BTREE_ITER_INTENT);
+	k = bch2_btree_iter_peek_with_updates(iter);
+
+	while (k.k && !(ret = bkey_err(k))) {
+		if (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0)
 			break;
-		default:
-			BUG();
+
+		if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
+			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				break;
+
+			bkey_reassemble(update, k);
+
+			bch2_cut_back(start, update);
+
+			update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
+							  BTREE_ITER_NOT_EXTENTS|
+							  BTREE_ITER_INTENT);
+			bch2_trans_update2(trans, update_iter, update);
+			bch2_trans_iter_put(trans, update_iter);
 		}
 
-		if (!trans->did_work && (ret || split))
+		if (bkey_cmp(k.k->p, insert->k.p) < 0 ||
+		    (!bkey_cmp(k.k->p, insert->k.p) && bkey_deleted(&insert->k))) {
+			update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				break;
+
+			bkey_init(&update->k);
+			update->k.p = k.k->p;
+
+			update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
+							  BTREE_ITER_NOT_EXTENTS|
+							  BTREE_ITER_INTENT);
+			bch2_trans_update2(trans, update_iter, update);
+			bch2_trans_iter_put(trans, update_iter);
+		}
+
+		if (bkey_cmp(k.k->p, insert->k.p) > 0) {
+			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				break;
+
+			bkey_reassemble(update, k);
+			bch2_cut_front(insert->k.p, update);
+
+			update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
+							  BTREE_ITER_NOT_EXTENTS|
+							  BTREE_ITER_INTENT);
+			bch2_trans_update2(trans, update_iter, update);
+			bch2_trans_iter_put(trans, update_iter);
 			break;
+		}
+
+		k = bch2_btree_iter_next_with_updates(iter);
 	}
-unlock:
-	multi_unlock_write(trans);
-	bch2_journal_res_put(&c->journal, &trans->journal_res);
+	bch2_trans_iter_put(trans, iter);
 
-	if (split)
-		goto split;
-	if (ret)
-		goto err;
+	return ret;
+}
 
-	trans_for_each_entry(trans, i)
-		if (i->iter->flags & BTREE_ITER_AT_END_OF_LEAF)
-			goto out;
+int __bch2_trans_commit(struct btree_trans *trans)
+{
+	struct btree_insert_entry *i = NULL;
+	struct btree_iter *iter;
+	bool trans_trigger_run;
+	unsigned u64s, reset_flags = 0;
+	int ret = 0;
 
-	trans_for_each_entry(trans, i) {
-		/*
-		 * iterators are inconsistent when they hit end of leaf, until
-		 * traversed again
-		 */
-		if (i->iter->uptodate < BTREE_ITER_NEED_TRAVERSE &&
-		    !same_leaf_as_prev(trans, i))
-			bch2_foreground_maybe_merge(c, i->iter, 0);
+	if (!trans->nr_updates)
+		goto out_reset;
+
+	if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
+		lockdep_assert_held(&trans->c->gc_lock);
+
+	memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
+
+	trans->journal_u64s		= trans->extra_journal_entry_u64s;
+	trans->journal_preres_u64s	= 0;
+
+	if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
+	    unlikely(!percpu_ref_tryget(&trans->c->writes))) {
+		ret = bch2_trans_commit_get_rw_cold(trans);
+		if (ret)
+			goto out_reset;
 	}
-out:
-	/* make sure we didn't lose an error: */
-	if (!ret && IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
-		trans_for_each_entry(trans, i)
-			BUG_ON(!i->done);
 
-	percpu_ref_put(&c->writes);
-	return ret;
-split:
+#ifdef CONFIG_BCACHEFS_DEBUG
+	trans_for_each_update(trans, i)
+		if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
+		    !(i->trigger_flags & BTREE_TRIGGER_NORUN))
+			bch2_btree_key_cache_verify_clean(trans,
+					i->btree_id, i->k->k.p);
+#endif
+
 	/*
-	 * have to drop journal res before splitting, because splitting means
-	 * allocating new btree nodes, and holding a journal reservation
-	 * potentially blocks the allocator:
+	 * Running triggers will append more updates to the list of updates as
+	 * we're walking it:
 	 */
-	ret = bch2_btree_split_leaf(c, split, trans->flags);
+	do {
+		trans_trigger_run = false;
+
+		trans_for_each_update(trans, i) {
+			if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
+			    !i->trans_triggers_run) {
+				i->trans_triggers_run = true;
+				trans_trigger_run = true;
+
+				ret = bch2_trans_mark_update(trans, i->iter, i->k,
+							     i->trigger_flags);
+				if (unlikely(ret)) {
+					if (ret == -EINTR)
+						trace_trans_restart_mark(trans->ip);
+					goto out;
+				}
+			}
+		}
+	} while (trans_trigger_run);
+
+	/* Turn extents updates into keys: */
+	trans_for_each_update(trans, i)
+		if (i->is_extent) {
+			ret = extent_handle_overwrites(trans, i->btree_id, i->k);
+			if (unlikely(ret))
+				goto out;
+		}
+
+	trans_for_each_update(trans, i) {
+		ret = i->is_extent
+			? extent_update_to_keys(trans, *i)
+			: (__bch2_trans_update2(trans, *i), 0);
+		if (unlikely(ret))
+			goto out;
+	}
+
+	trans_for_each_update2(trans, i) {
+		ret = bch2_btree_iter_traverse(i->iter);
+		if (unlikely(ret)) {
+			trace_trans_restart_traverse(trans->ip);
+			goto out;
+		}
+
+		if (unlikely(!bch2_btree_iter_upgrade(i->iter, i->level + 1))) {
+			trace_trans_restart_upgrade(trans->ip);
+			ret = -EINTR;
+			goto out;
+		}
+
+		BUG_ON(!btree_node_intent_locked(i->iter, i->level));
+
+		u64s = jset_u64s(i->k->k.u64s);
+		if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
+		    likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
+			trans->journal_preres_u64s += u64s;
+		trans->journal_u64s += u64s;
+	}
+retry:
+	memset(&trans->journal_res, 0, sizeof(trans->journal_res));
+
+	ret = do_bch2_trans_commit(trans, &i);
+
+	/* make sure we didn't drop or screw up locks: */
+	bch2_btree_trans_verify_locks(trans);
+
 	if (ret)
 		goto err;
-	/*
-	 * if the split didn't have to drop locks the insert will still be
-	 * atomic (in the BTREE_INSERT_ATOMIC sense, what the caller peeked()
-	 * and is overwriting won't have changed)
-	 */
-	goto retry_locks;
+
+	trans_for_each_iter(trans, iter)
+		if (btree_iter_live(trans, iter) &&
+		    (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT))
+			bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
+out:
+	bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
+
+	if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
+		percpu_ref_put(&trans->c->writes);
+out_reset:
+	if (!ret)
+		reset_flags |= TRANS_RESET_NOTRAVERSE;
+	if (!ret && (trans->flags & BTREE_INSERT_NOUNLOCK))
+		reset_flags |= TRANS_RESET_NOUNLOCK;
+	bch2_trans_reset(trans, reset_flags);
+
+	return ret;
 err:
-	if (cycle_gc_lock) {
-		down_read(&c->gc_lock);
-		up_read(&c->gc_lock);
+	ret = bch2_trans_commit_error(trans, i, ret);
+	if (ret)
+		goto out;
+
+	goto retry;
+}
+
+int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
+		      struct bkey_i *k, enum btree_trigger_flags flags)
+{
+	struct btree_insert_entry *i, n = (struct btree_insert_entry) {
+		.trigger_flags	= flags,
+		.bkey_type	= __btree_node_type(iter->level, iter->btree_id),
+		.btree_id	= iter->btree_id,
+		.level		= iter->level,
+		.is_extent	= (iter->flags & BTREE_ITER_IS_EXTENTS) != 0,
+		.iter		= iter,
+		.k		= k
+	};
+
+	BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+	BUG_ON(bkey_cmp(iter->pos,
+			n.is_extent ? bkey_start_pos(&k->k) : k->k.p));
+
+	trans_for_each_update(trans, i) {
+		BUG_ON(bkey_cmp(i->iter->pos,
+				i->is_extent ? bkey_start_pos(&i->k->k) : i->k->k.p));
+
+		BUG_ON(i != trans->updates &&
+		       btree_insert_entry_cmp(i - 1, i) >= 0);
 	}
+#endif
 
-	if (ret == -EINTR) {
-		trans_for_each_entry(trans, i) {
-			int ret2 = bch2_btree_iter_traverse(i->iter);
-			if (ret2) {
-				ret = ret2;
-				goto out;
+	iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+
+	if (n.is_extent) {
+		iter->pos_after_commit = k->k.p;
+		iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
+	}
+
+	/*
+	 * Pending updates are kept sorted: first, find position of new update,
+	 * then delete/trim any updates the new update overwrites:
+	 */
+	if (!n.is_extent) {
+		trans_for_each_update(trans, i)
+			if (btree_insert_entry_cmp(&n, i) <= 0)
+				break;
+
+		if (i < trans->updates + trans->nr_updates &&
+		    !btree_insert_entry_cmp(&n, i))
+			*i = n;
+		else
+			array_insert_item(trans->updates, trans->nr_updates,
+					  i - trans->updates, n);
+	} else {
+		trans_for_each_update(trans, i)
+			if (btree_insert_entry_cmp(&n, i) < 0)
+				break;
+
+		while (i > trans->updates &&
+		       i[-1].btree_id == n.btree_id &&
+		       bkey_cmp(bkey_start_pos(&n.k->k),
+				bkey_start_pos(&i[-1].k->k)) <= 0) {
+			--i;
+			array_remove_item(trans->updates, trans->nr_updates,
+					  i - trans->updates);
+		}
+
+		if (i > trans->updates &&
+		    i[-1].btree_id == n.btree_id &&
+		    bkey_cmp(bkey_start_pos(&n.k->k), i[-1].k->k.p) < 0)
+			bch2_cut_back(bkey_start_pos(&n.k->k), i[-1].k);
+
+		if (i < trans->updates + trans->nr_updates &&
+		    i->btree_id == n.btree_id &&
+		    bkey_cmp(n.k->k.p, bkey_start_pos(&i->k->k)) > 0) {
+			/* We don't handle splitting extents here: */
+			BUG_ON(bkey_cmp(bkey_start_pos(&n.k->k),
+					bkey_start_pos(&i->k->k)) > 0);
+
+			/*
+			 * When we have an extent that overwrites the start of another
+			 * update, trimming that extent will mean the iterator's
+			 * position has to change since the iterator position has to
+			 * match the extent's start pos - but we don't want to change
+			 * the iterator pos if some other code is using it, so we may
+			 * need to clone it:
+			 */
+			if (btree_iter_live(trans, i->iter)) {
+				i->iter = bch2_trans_copy_iter(trans, i->iter);
+
+				i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+				bch2_trans_iter_put(trans, i->iter);
 			}
+
+			bch2_cut_front(n.k->k.p, i->k);
+			bch2_btree_iter_set_pos(i->iter, n.k->k.p);
 		}
 
-		/*
-		 * BTREE_ITER_ATOMIC means we have to return -EINTR if we
-		 * dropped locks:
-		 */
-		if (!(trans->flags & BTREE_INSERT_ATOMIC))
-			goto retry;
+		array_insert_item(trans->updates, trans->nr_updates,
+				  i - trans->updates, n);
 	}
 
-	goto out;
+	return 0;
 }
 
-int bch2_btree_delete_at(struct btree_iter *iter, unsigned flags)
+void bch2_trans_commit_hook(struct btree_trans *trans,
+			    struct btree_trans_commit_hook *h)
 {
-	struct bkey_i k;
-
-	bkey_init(&k.k);
-	k.k.p = iter->pos;
-
-	return bch2_btree_insert_at(iter->c, NULL, NULL, NULL,
-				    BTREE_INSERT_NOFAIL|
-				    BTREE_INSERT_USE_RESERVE|flags,
-				    BTREE_INSERT_ENTRY(iter, &k));
+	h->next = trans->hooks;
+	trans->hooks = h;
 }
 
-int bch2_btree_insert_list_at(struct btree_iter *iter,
-			     struct keylist *keys,
-			     struct disk_reservation *disk_res,
-			     struct extent_insert_hook *hook,
-			     u64 *journal_seq, unsigned flags)
+int __bch2_btree_insert(struct btree_trans *trans,
+			enum btree_id id, struct bkey_i *k)
 {
-	BUG_ON(flags & BTREE_INSERT_ATOMIC);
-	BUG_ON(bch2_keylist_empty(keys));
-	bch2_verify_keylist_sorted(keys);
-
-	while (!bch2_keylist_empty(keys)) {
-		int ret = bch2_btree_insert_at(iter->c, disk_res, hook,
-				journal_seq, flags,
-				BTREE_INSERT_ENTRY(iter, bch2_keylist_front(keys)));
-		if (ret)
-			return ret;
+	struct btree_iter *iter;
+	int ret;
 
-		bch2_keylist_pop_front(keys);
-	}
+	iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
+				   BTREE_ITER_INTENT);
 
-	return 0;
+	ret = bch2_trans_update(trans, iter, k, 0);
+	bch2_trans_iter_put(trans, iter);
+	return ret;
 }
 
 /**
- * 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
  * @hook:		insert callback
  */
 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
-		     struct bkey_i *k,
-		     struct disk_reservation *disk_res,
-		     struct extent_insert_hook *hook,
-		     u64 *journal_seq, int flags)
+		      struct bkey_i *k,
+		      struct disk_reservation *disk_res,
+		      u64 *journal_seq, int flags)
 {
-	struct btree_iter iter;
-	int ret;
+	return bch2_trans_do(c, disk_res, journal_seq, flags,
+			     __bch2_btree_insert(&trans, id, k));
+}
 
-	bch2_btree_iter_init(&iter, c, id, bkey_start_pos(&k->k),
-			     BTREE_ITER_INTENT);
-	ret = bch2_btree_insert_at(c, disk_res, hook, journal_seq, flags,
-				  BTREE_INSERT_ENTRY(&iter, k));
-	bch2_btree_iter_unlock(&iter);
+int bch2_btree_delete_at(struct btree_trans *trans,
+			 struct btree_iter *iter, unsigned flags)
+{
+	struct bkey_i k;
 
-	return ret;
+	bkey_init(&k.k);
+	k.k.p = iter->pos;
+
+	bch2_trans_update(trans, iter, &k, 0);
+	return bch2_trans_commit(trans, NULL, NULL,
+				 BTREE_INSERT_NOFAIL|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,
-			   struct bversion version,
-			   struct disk_reservation *disk_res,
-			   struct extent_insert_hook *hook,
-			   u64 *journal_seq)
-{
-	struct btree_iter iter;
+int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
+				  struct bpos start, struct bpos end,
+				  u64 *journal_seq)
+{
+	struct btree_iter *iter;
 	struct bkey_s_c k;
 	int ret = 0;
 
-	bch2_btree_iter_init(&iter, c, id, start,
-			     BTREE_ITER_INTENT);
-
-	while ((k = bch2_btree_iter_peek(&iter)).k &&
-	       !(ret = btree_iter_err(k))) {
-		unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits);
-		/* really shouldn't be using a bare, unpadded bkey_i */
+	iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
+retry:
+	while ((k = bch2_btree_iter_peek(iter)).k &&
+	       !(ret = bkey_err(k)) &&
+	       bkey_cmp(iter->pos, end) < 0) {
 		struct bkey_i delete;
 
-		if (bkey_cmp(iter.pos, end) >= 0)
-			break;
+		bch2_trans_begin(trans);
 
 		bkey_init(&delete.k);
 
+		/*
+		 * This could probably be more efficient for extents:
+		 */
+
 		/*
 		 * For extents, iter.pos won't necessarily be the same as
 		 * bkey_start_pos(k.k) (for non extents they always will be the
@@ -579,33 +1212,48 @@ int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 		 * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 		 * bkey_start_pos(k.k)).
 		 */
-		delete.k.p = iter.pos;
-		delete.k.version = version;
+		delete.k.p = iter->pos;
 
-		if (iter.flags & BTREE_ITER_IS_EXTENTS) {
-			/*
-			 * The extents btree is special - KEY_TYPE_DISCARD is
-			 * used for deletions, not KEY_TYPE_DELETED. This is an
-			 * internal implementation detail that probably
-			 * shouldn't be exposed (internally, KEY_TYPE_DELETED is
-			 * used as a proxy for k->size == 0):
-			 */
-			delete.k.type = KEY_TYPE_DISCARD;
+		if (btree_node_type_is_extents(iter->btree_id)) {
+			unsigned max_sectors =
+				KEY_SIZE_MAX & (~0 << trans->c->block_bits);
 
 			/* create the biggest key we can */
 			bch2_key_resize(&delete.k, max_sectors);
-			bch2_cut_back(end, &delete.k);
+			bch2_cut_back(end, &delete);
+
+			ret = bch2_extent_trim_atomic(&delete, iter);
+			if (ret)
+				break;
 		}
 
-		ret = bch2_btree_insert_at(c, disk_res, hook, journal_seq,
-					  BTREE_INSERT_NOFAIL,
-					  BTREE_INSERT_ENTRY(&iter, &delete));
+		bch2_trans_update(trans, iter, &delete, 0);
+		ret = bch2_trans_commit(trans, NULL, journal_seq,
+					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;
+	}
+
+	bch2_trans_iter_free(trans, iter);
 	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)
+{
+	return bch2_trans_do(c, NULL, journal_seq, 0,
+			     bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
+}