X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_update_leaf.c;h=6cbc0e5e1fe819a703625fc7563823ee88201c6c;hb=284c1f5148afb088e64c8a77983a43732c4d499b;hp=b793ab77e452a21bf8a200d2bd9a63a4135bda0e;hpb=f87850496ea0ad0e5e8f8feb1c454662d0973530;p=bcachefs-tools-debian

diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_update_leaf.c
index b793ab7..6cbc0e5 100644
--- a/libbcachefs/btree_update_leaf.c
+++ b/libbcachefs/btree_update_leaf.c
@@ -10,56 +10,116 @@
 #include "btree_locking.h"
 #include "buckets.h"
 #include "debug.h"
+#include "errcode.h"
 #include "error.h"
 #include "extent_update.h"
 #include "journal.h"
 #include "journal_reclaim.h"
 #include "keylist.h"
+#include "recovery.h"
+#include "subvolume.h"
 #include "replicas.h"
 
 #include <linux/prefetch.h>
 #include <linux/sort.h>
 #include <trace/events/bcachefs.h>
 
+/*
+ * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
+ * different snapshot:
+ */
+struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
+{
+	struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
+
+	if (k.k && bpos_eq(path->pos, k.k->p))
+		return k;
+
+	bkey_init(u);
+	u->p = path->pos;
+	return (struct bkey_s_c) { u, NULL };
+}
+
+static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
+{
+#ifdef CONFIG_BCACHEFS_DEBUG
+	struct bch_fs *c = trans->c;
+	struct bkey u;
+	struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
+
+	if (unlikely(trans->journal_replay_not_finished)) {
+		struct bkey_i *j_k =
+			bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
+
+		if (j_k)
+			k = bkey_i_to_s_c(j_k);
+	}
+
+	i->old_k.needs_whiteout = k.k->needs_whiteout;
+
+	BUG_ON(memcmp(&i->old_k, k.k, sizeof(struct bkey)));
+	BUG_ON(i->old_v != k.v);
+#endif
+}
+
+static int __must_check
+bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
+			  struct bkey_i *, enum btree_update_flags);
+
 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->cached,	r->cached) ?:
 		 -cmp_int(l->level,	r->level) ?:
 		 bpos_cmp(l->k->k.p,	r->k->k.p);
 }
 
+static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
+{
+	return i->path->l + i->level;
+}
+
 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;
+	return i != trans->updates &&
+		insert_l(&i[0])->b == insert_l(&i[-1])->b;
+}
+
+static inline bool same_leaf_as_next(struct btree_trans *trans,
+				     struct btree_insert_entry *i)
+{
+	return i + 1 < trans->updates + trans->nr_updates &&
+		insert_l(&i[0])->b == insert_l(&i[1])->b;
 }
 
-inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
-					    struct btree_iter *iter)
+inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
+					   struct btree_path *path,
+					   struct btree *b)
 {
-	bch2_btree_node_lock_write(b, iter);
+	struct bch_fs *c = trans->c;
 
-	if (btree_iter_type(iter) == BTREE_ITER_CACHED)
+	if (path->cached)
 		return;
 
 	if (unlikely(btree_node_just_written(b)) &&
 	    bch2_btree_post_write_cleanup(c, b))
-		bch2_btree_iter_reinit_node(iter, b);
+		bch2_trans_node_reinit_iter(trans, 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);
+		bch2_btree_init_next(trans, b);
 }
 
 /* Inserting into a given leaf node (last stage of insert): */
 
 /* Handle overwrites and do insert, for non extents: */
-bool bch2_btree_bset_insert_key(struct btree_iter *iter,
+bool bch2_btree_bset_insert_key(struct btree_trans *trans,
+				struct btree_path *path,
 				struct btree *b,
 				struct btree_node_iter *node_iter,
 				struct bkey_i *insert)
@@ -70,11 +130,10 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 	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(bpos_cmp(insert->k.p, b->data->min_key) < 0);
-	EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
+	EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
+	EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 	EBUG_ON(insert->k.u64s >
-		bch_btree_keys_u64s_remaining(iter->trans->c, b));
-	EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
+		bch_btree_keys_u64s_remaining(trans->c, b));
 
 	k = bch2_btree_node_iter_peek_all(node_iter, b);
 	if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
@@ -93,7 +152,7 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 		k->type = KEY_TYPE_deleted;
 
 		if (k->needs_whiteout)
-			push_whiteout(iter->trans->c, b, insert->k.p);
+			push_whiteout(trans->c, b, insert->k.p);
 		k->needs_whiteout = false;
 
 		if (k >= btree_bset_last(b)->start) {
@@ -101,7 +160,7 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 			bch2_bset_delete(b, k, clobber_u64s);
 			goto fix_iter;
 		} else {
-			bch2_btree_iter_fix_key_modified(iter, b, k);
+			bch2_btree_path_fix_key_modified(trans, b, k);
 		}
 
 		return true;
@@ -119,7 +178,7 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 			clobber_u64s = k->u64s;
 			goto overwrite;
 		} else {
-			bch2_btree_iter_fix_key_modified(iter, b, k);
+			bch2_btree_path_fix_key_modified(trans, b, k);
 		}
 	}
 
@@ -129,7 +188,7 @@ overwrite:
 	new_u64s = k->u64s;
 fix_iter:
 	if (clobber_u64s != new_u64s)
-		bch2_btree_node_iter_fix(iter, b, node_iter, k,
+		bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 					 clobber_u64s, new_u64s);
 	return true;
 }
@@ -140,11 +199,32 @@ static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 	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]);
+	struct btree_trans trans;
+	unsigned long old, new, v;
+	unsigned idx = w - b->writes;
+
+	bch2_trans_init(&trans, c, 0, 0);
+
+	btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
+	v = READ_ONCE(b->flags);
+
+	do {
+		old = new = v;
+
+		if (!(old & (1 << BTREE_NODE_dirty)) ||
+		    !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
+		    w->journal.seq != seq)
+			break;
+
+		new &= ~BTREE_WRITE_TYPE_MASK;
+		new |= BTREE_WRITE_journal_reclaim;
+		new |= 1 << BTREE_NODE_need_write;
+	} while ((v = cmpxchg(&b->flags, old, new)) != old);
 
-	btree_node_lock_type(c, b, SIX_LOCK_read);
-	bch2_btree_node_write_cond(c, b,
-		(btree_current_write(b) == w && w->journal.seq == seq));
+	btree_node_write_if_need(c, b, SIX_LOCK_read);
 	six_unlock_read(&b->c.lock);
+
+	bch2_trans_exit(&trans);
 	return 0;
 }
 
@@ -172,24 +252,20 @@ inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 /**
  * btree_insert_key - insert a key one key into a leaf node
  */
-static bool btree_insert_key_leaf(struct btree_trans *trans,
-				  struct btree_iter *iter,
-				  struct bkey_i *insert)
+static void btree_insert_key_leaf(struct btree_trans *trans,
+				  struct btree_insert_entry *insert)
 {
 	struct bch_fs *c = trans->c;
-	struct btree *b = iter_l(iter)->b;
+	struct btree *b = insert_l(insert)->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;
 
-	EBUG_ON(!iter->level &&
-		!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
-
-	if (unlikely(!bch2_btree_bset_insert_key(iter, b,
-					&iter_l(iter)->iter, insert)))
-		return false;
+	if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
+					&insert_l(insert)->iter, insert->k)))
+		return;
 
 	i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 					 le64_to_cpu(i->journal_seq)));
@@ -197,7 +273,7 @@ static bool btree_insert_key_leaf(struct btree_trans *trans,
 	bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 
 	if (unlikely(!btree_node_dirty(b)))
-		set_btree_node_dirty(c, b);
+		set_btree_node_dirty_acct(c, b);
 
 	live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 	u64s_added = (int) bset_u64s(t) - old_u64s;
@@ -209,10 +285,7 @@ static bool btree_insert_key_leaf(struct btree_trans *trans,
 
 	if (u64s_added > live_u64s_added &&
 	    bch2_maybe_compact_whiteouts(c, b))
-		bch2_btree_iter_reinit_node(iter, b);
-
-	trace_btree_insert_key(c, b, insert);
-	return true;
+		bch2_trans_node_reinit_iter(trans, b);
 }
 
 /* Cached btree updates: */
@@ -222,13 +295,20 @@ static bool btree_insert_key_leaf(struct btree_trans *trans,
 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 					     struct btree_insert_entry *i)
 {
-	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);
+	BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
+	BUG_ON(i->cached	!= i->path->cached);
+	BUG_ON(i->level		!= i->path->level);
+	BUG_ON(i->btree_id	!= i->path->btree_id);
+	EBUG_ON(!i->level &&
+		!(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
+		test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
+		i->k->k.p.snapshot &&
+		bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 }
 
 static noinline int
-bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
+bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags,
+				   unsigned long trace_ip)
 {
 	struct bch_fs *c = trans->c;
 	int ret;
@@ -236,62 +316,70 @@ bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 	bch2_trans_unlock(trans);
 
 	ret = bch2_journal_preres_get(&c->journal,
-			&trans->journal_preres, u64s, 0);
+			&trans->journal_preres,
+			trans->journal_preres_u64s,
+			(flags & JOURNAL_WATERMARK_MASK));
 	if (ret)
 		return ret;
 
-	if (!bch2_trans_relock(trans)) {
-		trace_trans_restart_journal_preres_get(trans->ip);
-		return -EINTR;
+	ret = bch2_trans_relock(trans);
+	if (ret) {
+		trace_and_count(c, trans_restart_journal_preres_get, trans, trace_ip, 0);
+		return ret;
 	}
 
 	return 0;
 }
 
-static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
-					     unsigned flags)
+static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
+						      unsigned flags)
 {
-	struct bch_fs *c = trans->c;
-	int ret;
-
-	if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
-		flags |= JOURNAL_RES_GET_RESERVED;
+	return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
+				    trans->journal_u64s, flags);
+}
 
-	ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
-				   trans->journal_u64s, flags);
+#define JSET_ENTRY_LOG_U64s		4
 
-	return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
+static noinline void journal_transaction_name(struct btree_trans *trans)
+{
+	struct bch_fs *c = trans->c;
+	struct journal *j = &c->journal;
+	struct jset_entry *entry =
+		bch2_journal_add_entry(j, &trans->journal_res,
+				       BCH_JSET_ENTRY_log, 0, 0,
+				       JSET_ENTRY_LOG_U64s);
+	struct jset_entry_log *l =
+		container_of(entry, struct jset_entry_log, entry);
+
+	strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 }
 
-static enum btree_insert_ret
-btree_key_can_insert(struct btree_trans *trans,
-		     struct btree_iter *iter,
-		     unsigned u64s)
+static inline int btree_key_can_insert(struct btree_trans *trans,
+				       struct btree *b, 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 -BCH_ERR_btree_insert_btree_node_full;
 
-	return BTREE_INSERT_OK;
+	return 0;
 }
 
-static enum btree_insert_ret
-btree_key_can_insert_cached(struct btree_trans *trans,
-			    struct btree_iter *iter,
-			    unsigned u64s)
+static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
+				       struct btree_path *path, unsigned u64s)
 {
-	struct bkey_cached *ck = (void *) iter->l[0].b;
+	struct bch_fs *c = trans->c;
+	struct bkey_cached *ck = (void *) path->l[0].b;
+	struct btree_insert_entry *i;
 	unsigned new_u64s;
 	struct bkey_i *new_k;
 
-	BUG_ON(iter->level);
+	EBUG_ON(path->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;
+	    bch2_btree_key_cache_must_wait(c) &&
+	    !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
+		return -BCH_ERR_btree_insert_need_journal_reclaim;
 
 	/*
 	 * bch2_varint_decode can read past the end of the buffer by at most 7
@@ -300,75 +388,209 @@ btree_key_can_insert_cached(struct btree_trans *trans,
 	u64s += 1;
 
 	if (u64s <= ck->u64s)
-		return BTREE_INSERT_OK;
+		return 0;
 
 	new_u64s	= roundup_pow_of_two(u64s);
 	new_k		= krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
-	if (!new_k)
+	if (!new_k) {
+		bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
+			bch2_btree_ids[path->btree_id], new_u64s);
 		return -ENOMEM;
+	}
+
+	trans_for_each_update(trans, i)
+		if (i->old_v == &ck->k->v)
+			i->old_v = &new_k->v;
 
 	ck->u64s	= new_u64s;
 	ck->k		= new_k;
-	return BTREE_INSERT_OK;
+	return 0;
 }
 
-static inline void do_btree_insert_one(struct btree_trans *trans,
-				       struct btree_insert_entry *i)
+/* Triggers: */
+
+static int run_one_mem_trigger(struct btree_trans *trans,
+			       struct btree_insert_entry *i,
+			       unsigned flags)
 {
-	struct bch_fs *c = trans->c;
-	struct journal *j = &c->journal;
-	bool did_work;
+	struct bkey_s_c old = { &i->old_k, i->old_v };
+	struct bkey_i *new = i->k;
+	int ret;
 
-	EBUG_ON(trans->journal_res.ref !=
-		!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
+	verify_update_old_key(trans, i);
 
-	i->k->k.needs_whiteout = false;
+	if (unlikely(flags & BTREE_TRIGGER_NORUN))
+		return 0;
 
-	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 (!btree_node_type_needs_gc(i->btree_id))
+		return 0;
+
+	if (bch2_bkey_ops[old.k->type].atomic_trigger ==
+	    bch2_bkey_ops[i->k->k.type].atomic_trigger &&
+	    ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
+		ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
+				BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
+	} else {
+		struct bkey		_deleted = KEY(0, 0, 0);
+		struct bkey_s_c		deleted = (struct bkey_s_c) { &_deleted, NULL };
 
-	if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
-		bch2_journal_add_keys(j, &trans->journal_res,
-				      i->btree_id,
-				      i->level,
-				      i->k);
+		_deleted.p = i->path->pos;
 
-		bch2_journal_set_has_inode(j, &trans->journal_res,
-					   i->k->k.p.inode);
+		ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
+				BTREE_TRIGGER_INSERT|flags) ?:
+			bch2_mark_key(trans, old, deleted,
+				BTREE_TRIGGER_OVERWRITE|flags);
+	}
 
-		if (trans->journal_seq)
-			*trans->journal_seq = trans->journal_res.seq;
+	return ret;
+}
+
+static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
+				 bool overwrite)
+{
+	/*
+	 * Transactional triggers create new btree_insert_entries, so we can't
+	 * pass them a pointer to a btree_insert_entry, that memory is going to
+	 * move:
+	 */
+	struct bkey old_k = i->old_k;
+	struct bkey_s_c old = { &old_k, i->old_v };
+
+	verify_update_old_key(trans, i);
+
+	if ((i->flags & BTREE_TRIGGER_NORUN) ||
+	    !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
+		return 0;
+
+	if (!i->insert_trigger_run &&
+	    !i->overwrite_trigger_run &&
+	    bch2_bkey_ops[old.k->type].trans_trigger ==
+	    bch2_bkey_ops[i->k->k.type].trans_trigger &&
+	    ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
+		i->overwrite_trigger_run = true;
+		i->insert_trigger_run = true;
+		return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
+					   BTREE_TRIGGER_INSERT|
+					   BTREE_TRIGGER_OVERWRITE|
+					   i->flags) ?: 1;
+	} else if (overwrite && !i->overwrite_trigger_run) {
+		i->overwrite_trigger_run = true;
+		return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
+	} else if (!overwrite && !i->insert_trigger_run) {
+		i->insert_trigger_run = true;
+		return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
+	} else {
+		return 0;
 	}
 }
 
-static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
+static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
+			      struct btree_insert_entry *btree_id_start)
 {
-	__bch2_btree_iter_unlock(iter);
+	struct btree_insert_entry *i;
+	bool trans_trigger_run;
+	int ret, overwrite;
+
+	for (overwrite = 1; overwrite >= 0; --overwrite) {
+
+		/*
+		 * Running triggers will append more updates to the list of updates as
+		 * we're walking it:
+		 */
+		do {
+			trans_trigger_run = false;
+
+			for (i = btree_id_start;
+			     i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
+			     i++) {
+				if (i->btree_id != btree_id)
+					continue;
+
+				ret = run_one_trans_trigger(trans, i, overwrite);
+				if (ret < 0)
+					return ret;
+				if (ret)
+					trans_trigger_run = true;
+			}
+		} while (trans_trigger_run);
+	}
+
+	return 0;
+}
+
+static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
+{
+	struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
+	unsigned btree_id = 0;
+	int ret = 0;
+
+	/*
+	 *
+	 * For a given btree, this algorithm runs insert triggers before
+	 * overwrite triggers: this is so that when extents are being moved
+	 * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
+	 * they are re-added.
+	 */
+	for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
+		if (btree_id == BTREE_ID_alloc)
+			continue;
+
+		while (btree_id_start < trans->updates + trans->nr_updates &&
+		       btree_id_start->btree_id < btree_id)
+			btree_id_start++;
+
+		ret = run_btree_triggers(trans, btree_id, btree_id_start);
+		if (ret)
+			return ret;
+	}
+
+	trans_for_each_update(trans, i) {
+		if (i->btree_id > BTREE_ID_alloc)
+			break;
+		if (i->btree_id == BTREE_ID_alloc) {
+			ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
+			if (ret)
+				return ret;
+			break;
+		}
+	}
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+	trans_for_each_update(trans, i)
+		BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
+		       (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
+		       (!i->insert_trigger_run || !i->overwrite_trigger_run));
+#endif
+	return 0;
 }
 
-static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
+static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
+	int ret = 0;
 
 	trans_for_each_update(trans, i) {
 		/*
 		 * XXX: synchronization of cached update triggers with gc
+		 * XXX: synchronization of interior node updates with gc
 		 */
-		BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
+		BUG_ON(i->cached || i->level);
 
-		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);
+		if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
+			ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
+			if (ret)
+				break;
+		}
 	}
+
+	return ret;
 }
 
 static inline int
-bch2_trans_commit_write_locked(struct btree_trans *trans,
-			       struct btree_insert_entry **stopped_at)
+bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
+			       struct btree_insert_entry **stopped_at,
+			       unsigned long trace_ip)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
@@ -378,8 +600,8 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 	int ret;
 
 	if (race_fault()) {
-		trace_trans_restart_fault_inject(trans->ip);
-		return -EINTR;
+		trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
+		return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 	}
 
 	/*
@@ -398,15 +620,15 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 		h = h->next;
 	}
 
-	trans_for_each_update2(trans, i) {
+	trans_for_each_update(trans, i) {
 		/* Multiple inserts might go to same leaf: */
 		if (!same_leaf_as_prev(trans, i))
 			u64s = 0;
 
 		u64s += i->k->k.u64s;
-		ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
-			? btree_key_can_insert(trans, i->iter, u64s)
-			: btree_key_can_insert_cached(trans, i->iter, u64s);
+		ret = !i->cached
+			? btree_key_can_insert(trans, insert_l(i)->b, u64s)
+			: btree_key_can_insert_cached(trans, flags, i->path, u64s);
 		if (ret) {
 			*stopped_at = i;
 			return ret;
@@ -416,202 +638,244 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 			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))) {
+	if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 		ret = bch2_trans_journal_res_get(trans,
+				(flags & JOURNAL_WATERMARK_MASK)|
 				JOURNAL_RES_GET_NONBLOCK);
 		if (ret)
-			goto err;
+			return ret;
+
+		if (unlikely(trans->journal_transaction_names))
+			journal_transaction_name(trans);
 	} 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 (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
+	    !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 		if (bch2_journal_seq_verify)
-			trans_for_each_update2(trans, i)
+			trans_for_each_update(trans, i)
 				i->k->k.version.lo = trans->journal_res.seq;
 		else if (bch2_inject_invalid_keys)
-			trans_for_each_update2(trans, i)
+			trans_for_each_update(trans, i)
 				i->k->k.version = MAX_VERSION;
 	}
 
+	if (trans->fs_usage_deltas &&
+	    bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
+		return -BCH_ERR_btree_insert_need_mark_replicas;
+
 	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 (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
+			ret = run_one_mem_trigger(trans, i, i->flags);
+			if (ret)
+				return ret;
+		}
 
-	if (marking && trans->fs_usage_deltas)
-		bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas);
+	if (unlikely(c->gc_pos.phase)) {
+		ret = bch2_trans_commit_run_gc_triggers(trans);
+		if  (ret)
+			return ret;
+	}
 
-	if (unlikely(c->gc_pos.phase))
-		bch2_trans_mark_gc(trans);
+	if (unlikely(trans->extra_journal_entries.nr)) {
+		memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
+				  trans->extra_journal_entries.data,
+				  trans->extra_journal_entries.nr);
 
-	trans_for_each_update2(trans, i)
-		do_btree_insert_one(trans, i);
-err:
-	if (marking) {
-		percpu_up_read(&c->mark_lock);
+		trans->journal_res.offset	+= trans->extra_journal_entries.nr;
+		trans->journal_res.u64s		-= trans->extra_journal_entries.nr;
+	}
+
+	if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
+		trans_for_each_update(trans, i) {
+			struct journal *j = &c->journal;
+			struct jset_entry *entry;
+
+			if (i->key_cache_already_flushed)
+				continue;
+
+			if (i->flags & BTREE_UPDATE_NOJOURNAL)
+				continue;
+
+			verify_update_old_key(trans, i);
+
+			if (trans->journal_transaction_names) {
+				entry = bch2_journal_add_entry(j, &trans->journal_res,
+						       BCH_JSET_ENTRY_overwrite,
+						       i->btree_id, i->level,
+						       i->old_k.u64s);
+				bkey_reassemble(&entry->start[0],
+						(struct bkey_s_c) { &i->old_k, i->old_v });
+			}
+
+			entry = bch2_journal_add_entry(j, &trans->journal_res,
+					       BCH_JSET_ENTRY_btree_keys,
+					       i->btree_id, i->level,
+					       i->k->k.u64s);
+			bkey_copy(&entry->start[0], i->k);
+		}
+
+		if (trans->journal_seq)
+			*trans->journal_seq = trans->journal_res.seq;
+	}
+
+	trans_for_each_update(trans, i) {
+		i->k->k.needs_whiteout = false;
+
+		if (!i->cached)
+			btree_insert_key_leaf(trans, i);
+		else if (!i->key_cache_already_flushed)
+			bch2_btree_insert_key_cached(trans, flags, i->path, i->k);
+		else {
+			bch2_btree_key_cache_drop(trans, i->path);
+			btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
+		}
 	}
 
 	return ret;
 }
 
-static noinline int maybe_do_btree_merge(struct btree_trans *trans, struct btree_iter *iter)
+static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 {
-	struct btree_insert_entry *i;
-	struct btree *b = iter_l(iter)->b;
-	struct bkey_s_c old;
-	int u64s_delta = 0;
-	int ret;
+	while (--i >= trans->updates) {
+		if (same_leaf_as_prev(trans, i))
+			continue;
 
-	/*
-	 * 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;
+		bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
+	}
 
-	BUG_ON(iter->level);
+	trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
+	return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
+}
+
+static inline int trans_lock_write(struct btree_trans *trans)
+{
+	struct btree_insert_entry *i;
 
-	trans_for_each_update2(trans, i) {
-		if (iter_l(i->iter)->b != b)
+	trans_for_each_update(trans, i) {
+		if (same_leaf_as_prev(trans, i))
 			continue;
 
-		old = bch2_btree_iter_peek_slot(i->iter);
-		ret = bkey_err(old);
-		if (ret)
-			return ret;
+		if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
+			return trans_lock_write_fail(trans, i);
 
-		u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
-		u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
+		bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 	}
 
-	return u64s_delta <= 0
-		? (bch2_foreground_maybe_merge(trans->c, iter, iter->level,
-				trans->flags & ~BTREE_INSERT_NOUNLOCK) ?: -EINTR)
-		: 0;
+	return 0;
+}
+
+static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
+{
+	struct btree_insert_entry *i;
+
+	trans_for_each_update(trans, i)
+		bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
+}
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
+						   struct btree_insert_entry *i,
+						   struct printbuf *err)
+{
+	struct bch_fs *c = trans->c;
+	int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
+
+	printbuf_reset(err);
+	prt_printf(err, "invalid bkey on insert from %s -> %ps",
+		   trans->fn, (void *) i->ip_allocated);
+	prt_newline(err);
+	printbuf_indent_add(err, 2);
+
+	bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
+	prt_newline(err);
+
+	bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
+			  i->bkey_type, rw, err);
+	bch2_print_string_as_lines(KERN_ERR, err->buf);
+
+	bch2_inconsistent_error(c);
+	bch2_dump_trans_updates(trans);
+	printbuf_exit(err);
+
+	return -EINVAL;
 }
+#endif
 
 /*
  * Get journal reservation, take write locks, and attempt to do btree update(s):
  */
-static inline int do_bch2_trans_commit(struct btree_trans *trans,
-				       struct btree_insert_entry **stopped_at)
+static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
+				       struct btree_insert_entry **stopped_at,
+				       unsigned long trace_ip)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
-	struct btree_iter *iter;
-	int ret;
+	struct printbuf buf = PRINTBUF;
+	int ret, u64s_delta = 0;
 
-	trans_for_each_update2(trans, i) {
-		struct btree *b;
+#ifdef CONFIG_BCACHEFS_DEBUG
+	trans_for_each_update(trans, i) {
+		int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 
-		BUG_ON(!btree_node_intent_locked(i->iter, i->level));
+		if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
+					       i->bkey_type, rw, &buf)))
+			return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
+		btree_insert_entry_checks(trans, i);
+	}
+#endif
+	printbuf_exit(&buf);
 
-		if (btree_iter_type(i->iter) == BTREE_ITER_CACHED)
+	trans_for_each_update(trans, i) {
+		if (i->cached)
 			continue;
 
-		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;
+		u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
+		u64s_delta -= i->old_btree_u64s;
+
+		if (!same_leaf_as_next(trans, i)) {
+			if (u64s_delta <= 0) {
+				ret = bch2_foreground_maybe_merge(trans, i->path,
+							i->level, flags);
+				if (unlikely(ret))
+					return ret;
+			}
+
+			u64s_delta = 0;
 		}
 	}
 
-	trans_for_each_update2(trans, i)
-		BUG_ON(!btree_node_intent_locked(i->iter, i->level));
-
 	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);
+			(flags & JOURNAL_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
+	if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
+		ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
 	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);
-			}
-		}
-	}
-
-	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];
-
-			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);
+	ret = trans_lock_write(trans);
+	if (unlikely(ret))
+		return ret;
 
-	trans_for_each_update2(trans, i)
-		if (!same_leaf_as_prev(trans, i))
-			bch2_btree_node_lock_for_insert(c,
-					iter_l(i->iter)->b, i->iter);
+	ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 
-	ret = bch2_trans_commit_write_locked(trans, stopped_at);
+	if (!ret && unlikely(trans->journal_replay_not_finished))
+		bch2_drop_overwrites_from_journal(trans);
 
-	trans_for_each_update2(trans, i)
+	trans_for_each_update(trans, i)
 		if (!same_leaf_as_prev(trans, i))
-			bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
-							     i->iter);
+			bch2_btree_node_unlock_write_inlined(trans, i->path,
+							insert_l(i)->b);
 
 	if (!ret && trans->journal_pin)
 		bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
@@ -642,489 +906,704 @@ static int journal_reclaim_wait_done(struct bch_fs *c)
 }
 
 static noinline
-int bch2_trans_commit_error(struct btree_trans *trans,
+int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 			    struct btree_insert_entry *i,
-			    int ret)
+			    int ret, unsigned long trace_ip)
 {
 	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);
-
-		/*
-		 * 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 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;
-		}
+	case -BCH_ERR_btree_insert_btree_node_full:
+		ret = bch2_btree_split_leaf(trans, i->path, flags);
+		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+			trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 		break;
-	case BTREE_INSERT_ENOSPC:
-		ret = -ENOSPC;
-		break;
-	case BTREE_INSERT_NEED_MARK_REPLICAS:
+	case -BCH_ERR_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;
+			break;
 
-		trace_trans_restart_mark_replicas(trans->ip);
-		ret = -EINTR;
+		ret = bch2_trans_relock(trans);
+		if (ret)
+			trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 		break;
-	case BTREE_INSERT_NEED_JOURNAL_RES:
+	case -BCH_ERR_journal_res_get_blocked:
 		bch2_trans_unlock(trans);
 
-		if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
-		    !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED))
-			return -EAGAIN;
+		if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
+		    !(flags & JOURNAL_WATERMARK_reserved)) {
+			ret = -BCH_ERR_journal_reclaim_would_deadlock;
+			break;
+		}
 
-		ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
+		ret = bch2_trans_journal_res_get(trans,
+					(flags & JOURNAL_WATERMARK_MASK)|
+					JOURNAL_RES_GET_CHECK);
 		if (ret)
-			return ret;
-
-		if (bch2_trans_relock(trans))
-			return 0;
+			break;
 
-		trace_trans_restart_journal_res_get(trans->ip);
-		ret = -EINTR;
+		ret = bch2_trans_relock(trans);
+		if (ret)
+			trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 		break;
-	case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
+	case -BCH_ERR_btree_insert_need_journal_reclaim:
 		bch2_trans_unlock(trans);
 
+		trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_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;
+			break;
 
-		trace_trans_restart_journal_reclaim(trans->ip);
-		ret = -EINTR;
+		ret = bch2_trans_relock(trans);
+		if (ret)
+			trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 		break;
 	default:
 		BUG_ON(ret >= 0);
 		break;
 	}
 
+	BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
+
+	bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
+				!(flags & BTREE_INSERT_NOWAIT) &&
+				(flags & BTREE_INSERT_NOFAIL), c,
+		"%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
+
 	return ret;
 }
 
 static noinline int
-bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
+bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 {
 	struct bch_fs *c = trans->c;
 	int ret;
 
-	if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
-		return -EROFS;
+	if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
+	    test_bit(BCH_FS_STARTED, &c->flags))
+		return -BCH_ERR_erofs_trans_commit;
 
 	bch2_trans_unlock(trans);
 
-	ret = bch2_fs_read_write_early(c);
+	ret =   bch2_fs_read_write_early(c) ?:
+		bch2_trans_relock(trans);
 	if (ret)
 		return ret;
 
-	percpu_ref_get(&c->writes);
+	bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 	return 0;
 }
 
-static void __bch2_trans_update2(struct btree_trans *trans,
-				 struct btree_insert_entry n)
-{
+/*
+ * This is for updates done in the early part of fsck - btree_gc - before we've
+ * gone RW. we only add the new key to the list of keys for journal replay to
+ * do.
+ */
+static noinline int
+do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
+{
+	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
+	int ret = 0;
 
-	btree_insert_entry_checks(trans, &n);
+	trans_for_each_update(trans, i) {
+		ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
+		if (ret)
+			break;
+	}
 
-	EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX);
+	return ret;
+}
+
+int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_insert_entry *i = NULL;
+	unsigned u64s;
+	int ret = 0;
+
+	if (!trans->nr_updates &&
+	    !trans->extra_journal_entries.nr)
+		goto out_reset;
 
-	n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+	if (flags & BTREE_INSERT_GC_LOCK_HELD)
+		lockdep_assert_held(&c->gc_lock);
+
+	ret = bch2_trans_commit_run_triggers(trans);
+	if (ret)
+		goto out_reset;
+
+	if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
+		ret = do_bch2_trans_commit_to_journal_replay(trans);
+		goto out_reset;
+	}
+
+	if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
+	    unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
+		ret = bch2_trans_commit_get_rw_cold(trans, flags);
+		if (ret)
+			goto out_reset;
+	}
+
+	EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
+
+	memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
+
+	trans->journal_u64s		= trans->extra_journal_entries.nr;
+	trans->journal_preres_u64s	= 0;
+
+	trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
+
+	if (trans->journal_transaction_names)
+		trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
+
+	trans_for_each_update(trans, i) {
+		EBUG_ON(!i->path->should_be_locked);
+
+		ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
+		if (unlikely(ret))
+			goto out;
 
-	trans_for_each_update2(trans, i)
-		if (btree_insert_entry_cmp(&n, i) <= 0)
+		EBUG_ON(!btree_node_intent_locked(i->path, i->level));
+
+		if (i->key_cache_already_flushed)
+			continue;
+
+		/* we're going to journal the key being updated: */
+		u64s = jset_u64s(i->k->k.u64s);
+		if (i->cached &&
+		    likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
+			trans->journal_preres_u64s += u64s;
+
+		if (i->flags & BTREE_UPDATE_NOJOURNAL)
+			continue;
+
+		trans->journal_u64s += u64s;
+
+		/* and we're also going to log the overwrite: */
+		if (trans->journal_transaction_names)
+			trans->journal_u64s += jset_u64s(i->old_k.u64s);
+	}
+
+	if (trans->extra_journal_res) {
+		ret = bch2_disk_reservation_add(c, trans->disk_res,
+				trans->extra_journal_res,
+				(flags & BTREE_INSERT_NOFAIL)
+				? BCH_DISK_RESERVATION_NOFAIL : 0);
+		if (ret)
+			goto err;
+	}
+retry:
+	bch2_trans_verify_not_in_restart(trans);
+	memset(&trans->journal_res, 0, sizeof(trans->journal_res));
+
+	ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
+
+	/* make sure we didn't drop or screw up locks: */
+	bch2_trans_verify_locks(trans);
+
+	if (ret)
+		goto err;
+
+	trace_and_count(c, transaction_commit, trans, _RET_IP_);
+out:
+	bch2_journal_preres_put(&c->journal, &trans->journal_preres);
+
+	if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
+		bch2_write_ref_put(c, BCH_WRITE_REF_trans);
+out_reset:
+	bch2_trans_reset_updates(trans);
+
+	return ret;
+err:
+	ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
+	if (ret)
+		goto out;
+
+	goto retry;
+}
+
+static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
+					  enum btree_id id,
+					  struct bpos pos)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	bch2_trans_iter_init(trans, &iter, id, pos,
+			     BTREE_ITER_NOT_EXTENTS|
+			     BTREE_ITER_ALL_SNAPSHOTS);
+	while (1) {
+		k = bch2_btree_iter_prev(&iter);
+		ret = bkey_err(k);
+		if (ret)
+			break;
+
+		if (!k.k)
+			break;
+
+		if (!bkey_eq(pos, k.k->p))
+			break;
+
+		if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
+			ret = 1;
 			break;
+		}
+	}
+	bch2_trans_iter_exit(trans, &iter);
 
-	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);
+	return ret;
 }
 
-static void bch2_trans_update2(struct btree_trans *trans,
-			       struct btree_iter *iter,
-			       struct bkey_i *insert)
+static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
+					  enum btree_id id,
+					  struct bpos pos)
 {
-	__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,
-	});
+	if (!btree_type_has_snapshots(id) ||
+	    pos.snapshot == U32_MAX ||
+	    !snapshot_t(trans->c, pos.snapshot)->children[0])
+		return 0;
+
+	return __check_pos_snapshot_overwritten(trans, id, pos);
 }
 
-static int extent_update_to_keys(struct btree_trans *trans,
-				 struct btree_insert_entry n)
+static noinline int extent_front_merge(struct btree_trans *trans,
+				       struct btree_iter *iter,
+				       struct bkey_s_c k,
+				       struct bkey_i **insert,
+				       enum btree_update_flags flags)
 {
+	struct bch_fs *c = trans->c;
+	struct bkey_i *update;
 	int ret;
 
-	if (bkey_deleted(&n.k->k))
+	update = bch2_bkey_make_mut(trans, k);
+	ret = PTR_ERR_OR_ZERO(update);
+	if (ret)
+		return ret;
+
+	if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
+		return 0;
+
+	ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
+		check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
+	if (ret < 0)
+		return ret;
+	if (ret)
 		return 0;
 
-	ret = bch2_extent_can_insert(trans, n.iter, n.k);
+	ret = bch2_btree_delete_at(trans, iter, flags);
 	if (ret)
 		return ret;
 
-	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;
+	*insert = update;
+	return 0;
+}
+
+static noinline int extent_back_merge(struct btree_trans *trans,
+				      struct btree_iter *iter,
+				      struct bkey_i *insert,
+				      struct bkey_s_c k)
+{
+	struct bch_fs *c = trans->c;
+	int ret;
+
+	ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
+		check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
+	if (ret < 0)
+		return ret;
+	if (ret)
+		return 0;
 
-	__bch2_trans_update2(trans, n);
-	bch2_trans_iter_put(trans, n.iter);
+	bch2_bkey_merge(c, bkey_i_to_s(insert), k);
 	return 0;
 }
 
-static int extent_handle_overwrites(struct btree_trans *trans,
-				    enum btree_id btree_id,
-				    struct bkey_i *insert)
+int bch2_trans_update_extent(struct btree_trans *trans,
+			     struct btree_iter *orig_iter,
+			     struct bkey_i *insert,
+			     enum btree_update_flags flags)
 {
-	struct btree_iter *iter, *update_iter;
+	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;
+	enum btree_id btree_id = orig_iter->btree_id;
+	int ret = 0, compressed_sectors;
+
+	bch2_trans_iter_init(trans, &iter, btree_id, start,
+			     BTREE_ITER_INTENT|
+			     BTREE_ITER_WITH_UPDATES|
+			     BTREE_ITER_NOT_EXTENTS);
+	k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
+	if ((ret = bkey_err(k)))
+		goto err;
+	if (!k.k)
+		goto out;
 
-	iter = bch2_trans_get_iter(trans, btree_id, start,
-				   BTREE_ITER_INTENT);
-	k = bch2_btree_iter_peek_with_updates(iter);
+	if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
+		if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
+			ret = extent_front_merge(trans, &iter, k, &insert, flags);
+			if (ret)
+				goto err;
+		}
 
-	while (k.k && !(ret = bkey_err(k))) {
-		if (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0)
-			break;
+		goto next;
+	}
 
-		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;
+	while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
+		bool front_split = bkey_lt(bkey_start_pos(k.k), start);
+		bool back_split  = bkey_gt(k.k->p, insert->k.p);
 
-			bkey_reassemble(update, k);
+		/*
+		 * If we're going to be splitting a compressed extent, note it
+		 * so that __bch2_trans_commit() can increase our disk
+		 * reservation:
+		 */
+		if (((front_split && back_split) ||
+		     ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
+		    (compressed_sectors = bch2_bkey_sectors_compressed(k)))
+			trans->extra_journal_res += compressed_sectors;
+
+		if (front_split) {
+			update = bch2_bkey_make_mut(trans, k);
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				goto err;
 
 			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);
+			bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
+					     BTREE_ITER_NOT_EXTENTS|
+					     BTREE_ITER_ALL_SNAPSHOTS|
+					     BTREE_ITER_INTENT);
+			ret   = bch2_btree_iter_traverse(&update_iter) ?:
+				bch2_trans_update(trans, &update_iter, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+						  flags);
+			bch2_trans_iter_exit(trans, &update_iter);
+
+			if (ret)
+				goto err;
 		}
 
-		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 (k.k->p.snapshot != insert->k.p.snapshot &&
+		    (front_split || back_split)) {
+			update = bch2_bkey_make_mut(trans, k);
 			if ((ret = PTR_ERR_OR_ZERO(update)))
-				break;
+				goto err;
+
+			bch2_cut_front(start, update);
+			bch2_cut_back(insert->k.p, update);
+
+			bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
+					     BTREE_ITER_NOT_EXTENTS|
+					     BTREE_ITER_ALL_SNAPSHOTS|
+					     BTREE_ITER_INTENT);
+			ret   = bch2_btree_iter_traverse(&update_iter) ?:
+				bch2_trans_update(trans, &update_iter, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+						  flags);
+			bch2_trans_iter_exit(trans, &update_iter);
+			if (ret)
+				goto err;
+		}
+
+		if (bkey_le(k.k->p, insert->k.p)) {
+			update = bch2_trans_kmalloc(trans, sizeof(*update));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				goto err;
 
 			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 (insert->k.p.snapshot != k.k->p.snapshot) {
+				update->k.p.snapshot = insert->k.p.snapshot;
+				update->k.type = KEY_TYPE_whiteout;
+			}
+
+			bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
+					     BTREE_ITER_NOT_EXTENTS|
+					     BTREE_ITER_INTENT);
+			ret   = bch2_btree_iter_traverse(&update_iter) ?:
+				bch2_trans_update(trans, &update_iter, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+						  flags);
+			bch2_trans_iter_exit(trans, &update_iter);
+
+			if (ret)
+				goto err;
 		}
 
-		if (bkey_cmp(k.k->p, insert->k.p) > 0) {
-			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+		if (back_split) {
+			update = bch2_bkey_make_mut(trans, k);
 			if ((ret = PTR_ERR_OR_ZERO(update)))
-				break;
+				goto err;
 
-			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;
+			ret = bch2_trans_update_by_path(trans, iter.path, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+						  flags);
+			if (ret)
+				goto err;
+			goto out;
 		}
+next:
+		bch2_btree_iter_advance(&iter);
+		k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
+		if ((ret = bkey_err(k)))
+			goto err;
+		if (!k.k)
+			goto out;
+	}
 
-		k = bch2_btree_iter_next_with_updates(iter);
+	if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
+		ret = extent_back_merge(trans, &iter, insert, k);
+		if (ret)
+			goto err;
 	}
-	bch2_trans_iter_put(trans, iter);
+out:
+	if (!bkey_deleted(&insert->k)) {
+		/*
+		 * Rewinding iterators is expensive: get a new one and the one
+		 * that points to the start of insert will be cloned from:
+		 */
+		bch2_trans_iter_exit(trans, &iter);
+		bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
+				     BTREE_ITER_NOT_EXTENTS|
+				     BTREE_ITER_INTENT);
+		ret   = bch2_btree_iter_traverse(&iter) ?:
+			bch2_trans_update(trans, &iter, insert, flags);
+	}
+err:
+	bch2_trans_iter_exit(trans, &iter);
 
 	return ret;
 }
 
-int __bch2_trans_commit(struct btree_trans *trans)
+/*
+ * When deleting, check if we need to emit a whiteout (because we're overwriting
+ * something in an ancestor snapshot)
+ */
+static int need_whiteout_for_snapshot(struct btree_trans *trans,
+				      enum btree_id btree_id, struct bpos pos)
 {
-	struct btree_insert_entry *i = NULL;
-	struct btree_iter *iter;
-	bool trans_trigger_run;
-	unsigned u64s, reset_flags = 0;
-	int ret = 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;
-	}
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	u32 snapshot = pos.snapshot;
+	int ret;
 
-#ifdef CONFIG_BCACHEFS_DEBUG
-	trans_for_each_update(trans, i)
-		if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
-		    !(i->trigger_flags & BTREE_TRIGGER_NORUN))
-			bch2_btree_key_cache_verify_clean(trans,
-					i->btree_id, i->k->k.p);
-#endif
+	if (!bch2_snapshot_parent(trans->c, pos.snapshot))
+		return 0;
 
-	/*
-	 * Running triggers will append more updates to the list of updates as
-	 * we're walking it:
-	 */
-	do {
-		trans_trigger_run = false;
+	pos.snapshot++;
 
-		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);
+	for_each_btree_key_norestart(trans, iter, btree_id, pos,
+			   BTREE_ITER_ALL_SNAPSHOTS|
+			   BTREE_ITER_NOPRESERVE, k, ret) {
+		if (!bkey_eq(k.k->p, pos))
+			break;
 
-	/* 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;
+		if (bch2_snapshot_is_ancestor(trans->c, snapshot,
+					      k.k->p.snapshot)) {
+			ret = !bkey_whiteout(k.k);
+			break;
 		}
-
-	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;
 	}
+	bch2_trans_iter_exit(trans, &iter);
 
-	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;
-		}
+	return ret;
+}
 
-		BUG_ON(!btree_node_intent_locked(i->iter, i->level));
+static int __must_check
+bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
+				struct bkey_i *k, enum btree_update_flags flags,
+				unsigned long ip);
 
-		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));
+static noinline int flush_new_cached_update(struct btree_trans *trans,
+					    struct btree_path *path,
+					    struct btree_insert_entry *i,
+					    enum btree_update_flags flags,
+					    unsigned long ip)
+{
+	struct btree_path *btree_path;
+	int ret;
 
-	ret = do_bch2_trans_commit(trans, &i);
+	i->key_cache_already_flushed = true;
+	i->flags |= BTREE_TRIGGER_NORUN;
 
-	/* make sure we didn't drop or screw up locks: */
-	bch2_btree_trans_verify_locks(trans);
+	btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
+				   BTREE_ITER_INTENT, _THIS_IP_);
 
+	ret = bch2_btree_path_traverse(trans, btree_path, 0);
 	if (ret)
 		goto err;
 
-	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;
+	btree_path_set_should_be_locked(btree_path);
+	ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
 err:
-	ret = bch2_trans_commit_error(trans, i, ret);
-	if (ret)
-		goto out;
-
-	goto retry;
+	bch2_path_put(trans, btree_path, true);
+	return ret;
 }
 
-int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
-		      struct bkey_i *k, enum btree_trigger_flags flags)
+static int __must_check
+bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
+				struct bkey_i *k, enum btree_update_flags flags,
+				unsigned long ip)
 {
-	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
+	struct bch_fs *c = trans->c;
+	struct btree_insert_entry *i, n;
+	int cmp;
+
+	EBUG_ON(!path->should_be_locked);
+	EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
+	EBUG_ON(!bpos_eq(k->k.p, path->pos));
+
+	n = (struct btree_insert_entry) {
+		.flags		= flags,
+		.bkey_type	= __btree_node_type(path->level, path->btree_id),
+		.btree_id	= path->btree_id,
+		.level		= path->level,
+		.cached		= path->cached,
+		.path		= path,
+		.k		= k,
+		.ip_allocated	= ip,
 	};
 
-	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));
-
+	trans_for_each_update(trans, i)
 		BUG_ON(i != trans->updates &&
 		       btree_insert_entry_cmp(i - 1, i) >= 0);
-	}
 #endif
 
-	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;
+	trans_for_each_update(trans, i) {
+		cmp = btree_insert_entry_cmp(&n, i);
+		if (cmp <= 0)
+			break;
+	}
+
+	if (!cmp && i < trans->updates + trans->nr_updates) {
+		EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
 
-		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);
+		bch2_path_put(trans, i->path, true);
+		i->flags	= n.flags;
+		i->cached	= n.cached;
+		i->k		= n.k;
+		i->path		= n.path;
+		i->ip_allocated	= n.ip_allocated;
 	} else {
-		trans_for_each_update(trans, i)
-			if (btree_insert_entry_cmp(&n, i) < 0)
-				break;
+		array_insert_item(trans->updates, trans->nr_updates,
+				  i - trans->updates, n);
+
+		i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
+		i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
+
+		if (unlikely(trans->journal_replay_not_finished)) {
+			struct bkey_i *j_k =
+				bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
 
-		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 (j_k) {
+				i->old_k = j_k->k;
+				i->old_v = &j_k->v;
+			}
 		}
+	}
+
+	__btree_path_get(i->path, true);
+
+	/*
+	 * If a key is present in the key cache, it must also exist in the
+	 * btree - this is necessary for cache coherency. When iterating over
+	 * a btree that's cached in the key cache, the btree iter code checks
+	 * the key cache - but the key has to exist in the btree for that to
+	 * work:
+	 */
+	if (path->cached &&
+	    bkey_deleted(&i->old_k) &&
+	    !(flags & BTREE_UPDATE_NO_KEY_CACHE_COHERENCY))
+		return flush_new_cached_update(trans, path, i, flags, ip);
+
+	return 0;
+}
+
+static inline int __must_check
+bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
+			  struct bkey_i *k, enum btree_update_flags flags)
+{
+	return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
+}
+
+int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
+				   struct bkey_i *k, enum btree_update_flags flags)
+{
+	struct btree_path *path = iter->update_path ?: iter->path;
+	struct bkey_cached *ck;
+	int ret;
+
+	if (iter->flags & BTREE_ITER_IS_EXTENTS)
+		return bch2_trans_update_extent(trans, iter, k, flags);
+
+	if (bkey_deleted(&k->k) &&
+	    !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
+	    (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
+		ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
+		if (unlikely(ret < 0))
+			return ret;
+
+		if (ret)
+			k->k.type = KEY_TYPE_whiteout;
+	}
+
+	/*
+	 * Ensure that updates to cached btrees go to the key cache:
+	 */
+	if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
+	    !path->cached &&
+	    !path->level &&
+	    btree_id_cached(trans->c, path->btree_id)) {
+		if (!iter->key_cache_path ||
+		    !iter->key_cache_path->should_be_locked ||
+		    !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
+			if (!iter->key_cache_path)
+				iter->key_cache_path =
+					bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
+						      BTREE_ITER_INTENT|
+						      BTREE_ITER_CACHED, _THIS_IP_);
+
+			iter->key_cache_path =
+				bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
+							iter->flags & BTREE_ITER_INTENT,
+							_THIS_IP_);
+
+			ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
+						       BTREE_ITER_CACHED);
+			if (unlikely(ret))
+				return ret;
+
+			ck = (void *) iter->key_cache_path->l[0].b;
 
-		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);
+			if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+				trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
+				return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
 			}
 
-			bch2_cut_front(n.k->k.p, i->k);
-			bch2_btree_iter_set_pos(i->iter, n.k->k.p);
+			btree_path_set_should_be_locked(iter->key_cache_path);
 		}
 
-		array_insert_item(trans->updates, trans->nr_updates,
-				  i - trans->updates, n);
+		path = iter->key_cache_path;
 	}
 
-	return 0;
+	return bch2_trans_update_by_path(trans, path, k, flags);
 }
 
 void bch2_trans_commit_hook(struct btree_trans *trans,
@@ -1137,14 +1616,14 @@ void bch2_trans_commit_hook(struct btree_trans *trans,
 int __bch2_btree_insert(struct btree_trans *trans,
 			enum btree_id id, struct bkey_i *k)
 {
-	struct btree_iter *iter;
+	struct btree_iter iter;
 	int ret;
 
-	iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
-				   BTREE_ITER_INTENT);
-
-	ret = bch2_trans_update(trans, iter, k, 0);
-	bch2_trans_iter_put(trans, iter);
+	bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
+			     BTREE_ITER_INTENT);
+	ret   = bch2_btree_iter_traverse(&iter) ?:
+		bch2_trans_update(trans, &iter, k, 0);
+	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
 
@@ -1164,35 +1643,46 @@ int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
 			     __bch2_btree_insert(&trans, id, k));
 }
 
-int bch2_btree_delete_at(struct btree_trans *trans,
-			 struct btree_iter *iter, unsigned flags)
+int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
+				unsigned len, unsigned update_flags)
 {
-	struct bkey_i k;
+	struct bkey_i *k;
 
-	bkey_init(&k.k);
-	k.k.p = iter->pos;
+	k = bch2_trans_kmalloc(trans, sizeof(*k));
+	if (IS_ERR(k))
+		return PTR_ERR(k);
 
-	bch2_trans_update(trans, iter, &k, 0);
-	return bch2_trans_commit(trans, NULL, NULL,
-				 BTREE_INSERT_NOFAIL|flags);
+	bkey_init(&k->k);
+	k->k.p = iter->pos;
+	bch2_key_resize(&k->k, len);
+	return bch2_trans_update(trans, iter, k, update_flags);
+}
+
+int bch2_btree_delete_at(struct btree_trans *trans,
+			 struct btree_iter *iter, unsigned update_flags)
+{
+	return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
 }
 
 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
 				  struct bpos start, struct bpos end,
+				  unsigned update_flags,
 				  u64 *journal_seq)
 {
-	struct btree_iter *iter;
+	u32 restart_count = trans->restart_count;
+	struct btree_iter iter;
 	struct bkey_s_c k;
 	int ret = 0;
 
-	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) {
+	bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
+	while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
+		struct disk_reservation disk_res =
+			bch2_disk_reservation_init(trans->c, 0);
 		struct bkey_i delete;
 
-		bch2_trans_begin(trans);
+		ret = bkey_err(k);
+		if (ret)
+			goto err;
 
 		bkey_init(&delete.k);
 
@@ -1210,36 +1700,35 @@ retry:
 		 * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 		 * bkey_start_pos(k.k)).
 		 */
-		delete.k.p = iter->pos;
+		delete.k.p = iter.pos;
 
-		if (btree_node_type_is_extents(iter->btree_id)) {
-			unsigned max_sectors =
-				KEY_SIZE_MAX & (~0 << trans->c->block_bits);
+		if (iter.flags & BTREE_ITER_IS_EXTENTS)
+			bch2_key_resize(&delete.k,
+					bpos_min(end, k.k->p).offset -
+					iter.pos.offset);
 
-			/* create the biggest key we can */
-			bch2_key_resize(&delete.k, max_sectors);
-			bch2_cut_back(end, &delete);
-
-			ret = bch2_extent_trim_atomic(&delete, iter);
-			if (ret)
-				break;
-		}
+		ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
+			bch2_trans_commit(trans, &disk_res, journal_seq,
+					  BTREE_INSERT_NOFAIL);
+		bch2_disk_reservation_put(trans->c, &disk_res);
+err:
+		/*
+		 * the bch2_trans_begin() call is in a weird place because we
+		 * need to call it after every transaction commit, to avoid path
+		 * overflow, but don't want to call it if the delete operation
+		 * is a no-op and we have no work to do:
+		 */
+		bch2_trans_begin(trans);
 
-		bch2_trans_update(trans, iter, &delete, 0);
-		ret = bch2_trans_commit(trans, NULL, journal_seq,
-					BTREE_INSERT_NOFAIL);
+		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+			ret = 0;
 		if (ret)
 			break;
-
-		bch2_trans_cond_resched(trans);
-	}
-
-	if (ret == -EINTR) {
-		ret = 0;
-		goto retry;
 	}
+	bch2_trans_iter_exit(trans, &iter);
 
-	bch2_trans_iter_free(trans, iter);
+	if (!ret && trans_was_restarted(trans, restart_count))
+		ret = -BCH_ERR_transaction_restart_nested;
 	return ret;
 }
 
@@ -1250,8 +1739,81 @@ retry:
  */
 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 			    struct bpos start, struct bpos end,
+			    unsigned update_flags,
 			    u64 *journal_seq)
 {
-	return bch2_trans_do(c, NULL, journal_seq, 0,
-			     bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
+	int ret = bch2_trans_run(c,
+			bch2_btree_delete_range_trans(&trans, id, start, end,
+						      update_flags, journal_seq));
+	if (ret == -BCH_ERR_transaction_restart_nested)
+		ret = 0;
+	return ret;
+}
+
+static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
+{
+	struct printbuf buf = PRINTBUF;
+	struct jset_entry_log *l;
+	unsigned u64s;
+	int ret;
+
+	prt_vprintf(&buf, fmt, args);
+	ret = buf.allocation_failure ? -ENOMEM : 0;
+	if (ret)
+		goto err;
+
+	u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
+
+	ret = darray_make_room(entries, jset_u64s(u64s));
+	if (ret)
+		goto err;
+
+	l = (void *) &darray_top(*entries);
+	l->entry.u64s		= cpu_to_le16(u64s);
+	l->entry.btree_id	= 0;
+	l->entry.level		= 1;
+	l->entry.type		= BCH_JSET_ENTRY_log;
+	l->entry.pad[0]		= 0;
+	l->entry.pad[1]		= 0;
+	l->entry.pad[2]		= 0;
+	memcpy(l->d, buf.buf, buf.pos);
+	while (buf.pos & 7)
+		l->d[buf.pos++] = '\0';
+
+	entries->nr += jset_u64s(u64s);
+err:
+	printbuf_exit(&buf);
+	return ret;
+}
+
+int bch2_trans_log_msg(struct btree_trans *trans, const char *fmt, ...)
+{
+	va_list args;
+	int ret;
+
+	va_start(args, fmt);
+	ret = __bch2_trans_log_msg(&trans->extra_journal_entries, fmt, args);
+	va_end(args);
+
+	return ret;
+}
+
+int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
+{
+	va_list args;
+	int ret;
+
+	va_start(args, fmt);
+
+	if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
+		ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
+	} else {
+		ret = bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
+			__bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
+	}
+
+	va_end(args);
+
+	return ret;
+
 }