X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;ds=sidebyside;f=libbcachefs%2Fbtree_update_leaf.c;h=c93c132dd815d963360dfe5060dd74925e3ba915;hb=1ee7dc7a55273d34358a0ee525a9e823c999ffe6;hp=6f4ee55ebffe0f5214a19a68d5b74de0e638052b;hpb=e240b4ae86adb022e3266220ce9807dad8b51beb;p=bcachefs-tools-debian

diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_update_leaf.c
index 6f4ee55..c93c132 100644
--- a/libbcachefs/btree_update_leaf.c
+++ b/libbcachefs/btree_update_leaf.c
@@ -8,8 +8,10 @@
 #include "btree_iter.h"
 #include "btree_key_cache.h"
 #include "btree_locking.h"
+#include "btree_write_buffer.h"
 #include "buckets.h"
 #include "debug.h"
+#include "errcode.h"
 #include "error.h"
 #include "extent_update.h"
 #include "journal.h"
@@ -23,6 +25,45 @@
 #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);
+	}
+
+	u = *k.k;
+	u.needs_whiteout = i->old_k.needs_whiteout;
+
+	BUG_ON(memcmp(&i->old_k, &u, 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);
@@ -31,6 +72,7 @@ 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);
 }
@@ -54,15 +96,12 @@ static inline bool same_leaf_as_next(struct btree_trans *trans,
 		insert_l(&i[0])->b == insert_l(&i[1])->b;
 }
 
-static inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
-						  struct btree_path *path,
-						  struct btree *b)
+inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
+					   struct btree_path *path,
+					   struct btree *b)
 {
 	struct bch_fs *c = trans->c;
 
-	if (path->cached)
-		return;
-
 	if (unlikely(btree_node_just_written(b)) &&
 	    bch2_btree_post_write_cleanup(c, b))
 		bch2_trans_node_reinit_iter(trans, b);
@@ -75,14 +114,6 @@ static inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
 		bch2_btree_init_next(trans, b);
 }
 
-void bch2_btree_node_lock_for_insert(struct btree_trans *trans,
-				     struct btree_path *path,
-				     struct btree *b)
-{
-	bch2_btree_node_lock_write(trans, path, b);
-	bch2_btree_node_prep_for_write(trans, path, b);
-}
-
 /* Inserting into a given leaf node (last stage of insert): */
 
 /* Handle overwrites and do insert, for non extents: */
@@ -98,8 +129,8 @@ bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 	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(trans->c, b));
 
@@ -167,10 +198,13 @@ 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;
 
-	six_lock_read(&b->c.lock, NULL, NULL);
+	bch2_trans_init(&trans, c, 0, 0);
+
+	btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 	v = READ_ONCE(b->flags);
 
 	do {
@@ -181,11 +215,15 @@ static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 		    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_write_if_need(c, b, SIX_LOCK_read);
 	six_unlock_read(&b->c.lock);
+
+	bch2_trans_exit(&trans);
 	return 0;
 }
 
@@ -213,25 +251,26 @@ inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 /**
  * btree_insert_key - insert a key one key into a leaf node
  */
-static void btree_insert_key_leaf(struct btree_trans *trans,
-				  struct btree_insert_entry *insert)
+inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
+				       struct btree_path *path,
+				       struct bkey_i *insert,
+				       u64 journal_seq)
 {
 	struct bch_fs *c = trans->c;
-	struct btree *b = insert_l(insert)->b;
+	struct btree *b = path_l(path)->b;
 	struct bset_tree *t = bset_tree_last(b);
 	struct bset *i = bset(b, t);
 	int old_u64s = bset_u64s(t);
 	int old_live_u64s = b->nr.live_u64s;
 	int live_u64s_added, u64s_added;
 
-	if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
-					&insert_l(insert)->iter, insert->k)))
+	if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
+					&path_l(path)->iter, insert)))
 		return;
 
-	i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
-					 le64_to_cpu(i->journal_seq)));
+	i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 
-	bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
+	bch2_btree_add_journal_pin(c, b, journal_seq);
 
 	if (unlikely(!btree_node_dirty(b)))
 		set_btree_node_dirty_acct(c, b);
@@ -249,6 +288,12 @@ static void btree_insert_key_leaf(struct btree_trans *trans,
 		bch2_trans_node_reinit_iter(trans, b);
 }
 
+static void btree_insert_key_leaf(struct btree_trans *trans,
+				  struct btree_insert_entry *insert)
+{
+	bch2_btree_insert_key_leaf(trans, insert->path, insert->k, trans->journal_res.seq);
+}
+
 /* Cached btree updates: */
 
 /* Normal update interface: */
@@ -256,7 +301,7 @@ static void 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(bpos_cmp(i->k->k.p, i->path->pos));
+	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);
@@ -268,7 +313,7 @@ static inline void btree_insert_entry_checks(struct btree_trans *trans,
 }
 
 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;
@@ -277,30 +322,26 @@ 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->fn, trace_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;
-
-	ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
-				   trans->journal_u64s,
-				   flags|
-				   (trans->flags & JOURNAL_WATERMARK_MASK));
-
-	return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
+	return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
+				    trans->journal_u64s, flags);
 }
 
 #define JSET_ENTRY_LOG_U64s		4
@@ -308,56 +349,43 @@ static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 static noinline void journal_transaction_name(struct btree_trans *trans)
 {
 	struct bch_fs *c = trans->c;
-	struct jset_entry *entry = journal_res_entry(&c->journal, &trans->journal_res);
-	struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
-	unsigned u64s = JSET_ENTRY_LOG_U64s - 1;
-	unsigned b, buflen = u64s * sizeof(u64);
-
-	l->entry.u64s		= cpu_to_le16(u64s);
-	l->entry.btree_id	= 0;
-	l->entry.level		= 0;
-	l->entry.type		= BCH_JSET_ENTRY_log;
-	l->entry.pad[0]		= 0;
-	l->entry.pad[1]		= 0;
-	l->entry.pad[2]		= 0;
-	b = min_t(unsigned, strlen(trans->fn), buflen);
-	memcpy(l->d, trans->fn, b);
-	while (b < buflen)
-		l->d[b++] = '\0';
-
-	trans->journal_res.offset	+= JSET_ENTRY_LOG_U64s;
-	trans->journal_res.u64s		-= JSET_ENTRY_LOG_U64s;
+	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 inline enum btree_insert_ret
-btree_key_can_insert(struct btree_trans *trans,
-		     struct btree *b,
-		     unsigned u64s)
+static inline int btree_key_can_insert(struct btree_trans *trans,
+				       struct btree *b, unsigned u64s)
 {
 	struct bch_fs *c = trans->c;
 
 	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_path *path,
-			    unsigned u64s)
+static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
+				       struct btree_path *path, unsigned u64s)
 {
 	struct bch_fs *c = trans->c;
 	struct bkey_cached *ck = (void *) path->l[0].b;
-	unsigned old_u64s = ck->u64s, new_u64s;
+	struct btree_insert_entry *i;
+	unsigned new_u64s;
 	struct bkey_i *new_k;
 
 	EBUG_ON(path->level);
 
 	if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 	    bch2_btree_key_cache_must_wait(c) &&
-	    !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
-		return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
+	    !(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
@@ -366,7 +394,7 @@ 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);
@@ -376,48 +404,13 @@ btree_key_can_insert_cached(struct btree_trans *trans,
 		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;
-	/*
-	 * Keys returned by peek() are no longer valid pointers, so we need a
-	 * transaction restart:
-	 */
-	trace_trans_restart_key_cache_key_realloced(trans->fn, _RET_IP_,
-					     path->btree_id, &path->pos,
-					     old_u64s, new_u64s);
-	/*
-	 * Not using btree_trans_restart() because we can't unlock here, we have
-	 * write locks held:
-	 */
-	trans->restarted = true;
-	return -EINTR;
-}
-
-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;
-
-	EBUG_ON(trans->journal_res.ref !=
-		!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
-
-	i->k->k.needs_whiteout = false;
-
-	if (!i->cached)
-		btree_insert_key_leaf(trans, i);
-	else
-		bch2_btree_insert_key_cached(trans, i->path, i->k);
-
-	if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
-		bch2_journal_add_keys(j, &trans->journal_res,
-				      i->btree_id,
-				      i->level,
-				      i->k);
-
-		if (trans->journal_seq)
-			*trans->journal_seq = trans->journal_res.seq;
-	}
+	return 0;
 }
 
 /* Triggers: */
@@ -430,6 +423,8 @@ static int run_one_mem_trigger(struct btree_trans *trans,
 	struct bkey_i *new = i->k;
 	int ret;
 
+	verify_update_old_key(trans, i);
+
 	if (unlikely(flags & BTREE_TRIGGER_NORUN))
 		return 0;
 
@@ -439,7 +434,8 @@ static int run_one_mem_trigger(struct btree_trans *trans,
 	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),
+		ret   = bch2_mark_key(trans, i->btree_id, i->level,
+				old, bkey_i_to_s_c(new),
 				BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 	} else {
 		struct bkey		_deleted = KEY(0, 0, 0);
@@ -447,9 +443,11 @@ static int run_one_mem_trigger(struct btree_trans *trans,
 
 		_deleted.p = i->path->pos;
 
-		ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
+		ret   = bch2_mark_key(trans, i->btree_id, i->level,
+				deleted, bkey_i_to_s_c(new),
 				BTREE_TRIGGER_INSERT|flags) ?:
-			bch2_mark_key(trans, old, deleted,
+			bch2_mark_key(trans, i->btree_id, i->level,
+				old, deleted,
 				BTREE_TRIGGER_OVERWRITE|flags);
 	}
 
@@ -457,7 +455,7 @@ static int run_one_mem_trigger(struct btree_trans *trans,
 }
 
 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
-			   bool overwrite)
+				 bool overwrite)
 {
 	/*
 	 * Transactional triggers create new btree_insert_entries, so we can't
@@ -466,42 +464,33 @@ static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_
 	 */
 	struct bkey old_k = i->old_k;
 	struct bkey_s_c old = { &old_k, i->old_v };
-	int ret = 0;
+
+	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 (!overwrite) {
-		if (i->insert_trigger_run)
-			return 0;
-
-		BUG_ON(i->overwrite_trigger_run);
-		i->insert_trigger_run = true;
-	} else {
-		if (i->overwrite_trigger_run)
-			return 0;
-
-		BUG_ON(!i->insert_trigger_run);
-		i->overwrite_trigger_run = true;
-	}
-
-	if (overwrite) {
-		ret = bch2_trans_mark_old(trans, old, i->flags);
-	} else if (bch2_bkey_ops[old.k->type].trans_trigger ==
-		   bch2_bkey_ops[i->k->k.type].trans_trigger &&
+	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;
-		ret = bch2_trans_mark_key(trans, old, i->k,
-				BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|i->flags);
+		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 {
-		ret = bch2_trans_mark_new(trans, i->k, i->flags);
+		return 0;
 	}
-
-	if (ret == -EINTR)
-		trace_trans_restart_mark(trans->fn, _RET_IP_,
-					 i->btree_id, &i->path->pos);
-	return ret ?: 1;
 }
 
 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
@@ -511,7 +500,7 @@ static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 	bool trans_trigger_run;
 	int ret, overwrite;
 
-	for (overwrite = 0; overwrite < 2; overwrite++) {
+	for (overwrite = 1; overwrite >= 0; --overwrite) {
 
 		/*
 		 * Running triggers will append more updates to the list of updates as
@@ -575,11 +564,12 @@ static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 		}
 	}
 
+#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;
 }
 
@@ -607,21 +597,21 @@ static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 }
 
 static inline int
-bch2_trans_commit_write_locked(struct btree_trans *trans,
+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;
+	struct btree_write_buffered_key *wb;
 	struct btree_trans_commit_hook *h;
 	unsigned u64s = 0;
 	bool marking = false;
 	int ret;
 
 	if (race_fault()) {
-		trace_trans_restart_fault_inject(trans->fn, trace_ip);
-		trans->restarted = true;
-		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);
 	}
 
 	/*
@@ -648,7 +638,7 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 		u64s += i->k->k.u64s;
 		ret = !i->cached
 			? btree_key_can_insert(trans, insert_l(i)->b, u64s)
-			: btree_key_can_insert_cached(trans, i->path, u64s);
+			: btree_key_can_insert_cached(trans, flags, i->path, u64s);
 		if (ret) {
 			*stopped_at = i;
 			return ret;
@@ -656,40 +646,19 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 
 		if (btree_node_type_needs_gc(i->bkey_type))
 			marking = true;
-
-		/*
-		 * Revalidate before calling mem triggers - XXX, ugly:
-		 *
-		 * - successful btree node splits don't cause transaction
-		 *   restarts and will have invalidated the pointer to the bkey
-		 *   value
-		 * - btree_node_lock_for_insert() -> btree_node_prep_for_write()
-		 *   when it has to resort
-		 * - btree_key_can_insert_cached() when it has to reallocate
-		 *
-		 *   Ugly because we currently have no way to tell if the
-		 *   pointer's been invalidated, which means it's debatabale
-		 *   whether we should be stashing the old key at all.
-		 */
-		i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
-
-		if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
-			struct bkey_i *j_k =
-				bch2_journal_keys_peek(c, i->btree_id, i->level, i->k->k.p);
-
-			if (j_k && !bpos_cmp(j_k->k.p, i->k->k.p)) {
-				i->old_k = j_k->k;
-				i->old_v = &j_k->v;
-			}
-		}
 	}
 
+	if (trans->nr_wb_updates &&
+	    trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
+		return -BCH_ERR_btree_insert_need_flush_buffer;
+
 	/*
 	 * 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)
 			return ret;
@@ -700,21 +669,13 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 		trans->journal_res.seq = c->journal.replay_journal_seq;
 	}
 
-	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->journal_res.offset	+= trans->extra_journal_entries.nr;
-		trans->journal_res.u64s		-= trans->extra_journal_entries.nr;
-	}
-
 	/*
 	 * 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_update(trans, i)
 				i->k->k.version.lo = trans->journal_res.seq;
@@ -725,95 +686,112 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 
 	if (trans->fs_usage_deltas &&
 	    bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
-		return BTREE_INSERT_NEED_MARK_REPLICAS;
+		return -BCH_ERR_btree_insert_need_mark_replicas;
+
+	if (trans->nr_wb_updates) {
+		EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
+
+		ret = bch2_btree_insert_keys_write_buffer(trans);
+		if (ret)
+			goto revert_fs_usage;
+	}
 
 	trans_for_each_update(trans, i)
 		if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 			ret = run_one_mem_trigger(trans, i, i->flags);
 			if (ret)
-				return ret;
+				goto fatal_err;
 		}
 
 	if (unlikely(c->gc_pos.phase)) {
 		ret = bch2_trans_commit_run_gc_triggers(trans);
 		if  (ret)
-			return ret;
+			goto fatal_err;
 	}
 
-	trans_for_each_update(trans, i)
-		do_btree_insert_one(trans, i);
+	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);
 
-	return ret;
-}
+		trans->journal_res.offset	+= trans->extra_journal_entries.nr;
+		trans->journal_res.u64s		-= trans->extra_journal_entries.nr;
+	}
 
-static inline void path_upgrade_readers(struct btree_trans *trans, struct btree_path *path)
-{
-	unsigned l;
+	if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
+		struct journal *j = &c->journal;
+		struct jset_entry *entry;
 
-	for (l = 0; l < BTREE_MAX_DEPTH; l++)
-		if (btree_node_read_locked(path, l))
-			BUG_ON(!bch2_btree_node_upgrade(trans, path, l));
-}
+		trans_for_each_update(trans, i) {
+			if (i->key_cache_already_flushed)
+				continue;
 
-static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
-{
-	struct btree *b = path_l(path)->b;
+			if (i->flags & BTREE_UPDATE_NOJOURNAL)
+				continue;
 
-	do {
-		if (path->nodes_locked &&
-		    path->nodes_locked != path->nodes_intent_locked)
-			path_upgrade_readers(trans, path);
-	} while ((path = prev_btree_path(trans, path)) &&
-		 path_l(path)->b == b);
-}
+			verify_update_old_key(trans, i);
 
-/*
- * Check for nodes that we have both read and intent locks on, and upgrade the
- * readers to intent:
- */
-static inline void normalize_read_intent_locks(struct btree_trans *trans)
-{
-	struct btree_path *path;
-	unsigned i, nr_read = 0, nr_intent = 0;
+			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 });
+			}
 
-	trans_for_each_path_inorder(trans, path, i) {
-		struct btree_path *next = i + 1 < trans->nr_sorted
-			? trans->paths + trans->sorted[i + 1]
-			: NULL;
+			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 (path->nodes_locked) {
-			if (path->nodes_intent_locked)
-				nr_intent++;
-			else
-				nr_read++;
+		trans_for_each_wb_update(trans, wb) {
+			entry = bch2_journal_add_entry(j, &trans->journal_res,
+					       BCH_JSET_ENTRY_btree_keys,
+					       wb->btree, 0,
+					       wb->k.k.u64s);
+			bkey_copy(&entry->start[0], &wb->k);
 		}
 
-		if (!next || path_l(path)->b != path_l(next)->b) {
-			if (nr_read && nr_intent)
-				upgrade_readers(trans, path);
+		if (trans->journal_seq)
+			*trans->journal_seq = trans->journal_res.seq;
+	}
 
-			nr_read = nr_intent = 0;
+	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);
 		}
 	}
 
-	bch2_trans_verify_locks(trans);
+	return 0;
+fatal_err:
+	bch2_fatal_error(c);
+revert_fs_usage:
+	if (trans->fs_usage_deltas)
+		bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
+	return ret;
 }
 
-static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
+static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 {
-	struct btree_path *path;
-	unsigned i;
-
-	trans_for_each_path_inorder(trans, path, i) {
-		//if (path == pos)
-		//	break;
+	while (--i >= trans->updates) {
+		if (same_leaf_as_prev(trans, i))
+			continue;
 
-		if (path->nodes_locked != path->nodes_intent_locked &&
-		    !bch2_btree_path_upgrade(trans, path, path->level + 1))
-			return true;
+		bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 	}
 
-	return false;
+	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)
@@ -824,44 +802,61 @@ static inline int trans_lock_write(struct btree_trans *trans)
 		if (same_leaf_as_prev(trans, i))
 			continue;
 
-		if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
-			if (have_conflicting_read_lock(trans, i->path))
-				goto fail;
+		if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
+			return trans_lock_write_fail(trans, i);
 
-			btree_node_lock_type(trans, i->path,
-					     insert_l(i)->b,
-					     i->path->pos, i->level,
-					     SIX_LOCK_write, NULL, NULL);
-		}
-
-		bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
+		if (!i->cached)
+			bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 	}
 
 	return 0;
-fail:
-	while (--i >= trans->updates) {
-		if (same_leaf_as_prev(trans, i))
-			continue;
-
-		bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
-	}
-
-	trace_trans_restart_would_deadlock_write(trans->fn);
-	return btree_trans_restart(trans);
 }
 
 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 {
 	struct btree_insert_entry *i;
+	struct btree_write_buffered_key *wb;
 
 	trans_for_each_update(trans, i)
 		bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
+
+	trans_for_each_wb_update(trans, wb)
+		bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->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,
+static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 				       struct btree_insert_entry **stopped_at,
 				       unsigned long trace_ip)
 {
@@ -869,20 +864,19 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 	struct btree_insert_entry *i;
 	int ret, u64s_delta = 0;
 
+#ifdef CONFIG_BCACHEFS_DEBUG
+	struct printbuf buf = PRINTBUF;
+
 	trans_for_each_update(trans, i) {
-		const char *invalid = bch2_bkey_invalid(c,
-				bkey_i_to_s_c(i->k), i->bkey_type);
-		if (invalid) {
-			struct printbuf buf = PRINTBUF;
-
-			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
-			bch2_fs_fatal_error(c, "invalid bkey %s on insert from %s -> %ps: %s\n",
-					    buf.buf, trans->fn, (void *) i->ip_allocated, invalid);
-			printbuf_exit(&buf);
-			return -EINVAL;
-		}
+		int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
+
+		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);
 	}
+	printbuf_exit(&buf);
+#endif
 
 	trans_for_each_update(trans, i) {
 		if (i->cached)
@@ -894,7 +888,7 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 		if (!same_leaf_as_next(trans, i)) {
 			if (u64s_delta <= 0) {
 				ret = bch2_foreground_maybe_merge(trans, i->path,
-							i->level, trans->flags);
+							i->level, flags);
 				if (unlikely(ret))
 					return ret;
 			}
@@ -905,23 +899,19 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 
 	ret = bch2_journal_preres_get(&c->journal,
 			&trans->journal_preres, trans->journal_preres_u64s,
-			JOURNAL_RES_GET_NONBLOCK|
-			(trans->flags & JOURNAL_WATERMARK_MASK));
-	if (unlikely(ret == -EAGAIN))
-		ret = bch2_trans_journal_preres_get_cold(trans,
-						trans->journal_preres_u64s, trace_ip);
+			(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;
 
-	normalize_read_intent_locks(trans);
-
 	ret = trans_lock_write(trans);
 	if (unlikely(ret))
 		return ret;
 
-	ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
+	ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 
-	if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
+	if (!ret && unlikely(trans->journal_replay_not_finished))
 		bch2_drop_overwrites_from_journal(trans);
 
 	trans_for_each_update(trans, i)
@@ -958,104 +948,119 @@ 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, unsigned long trace_ip)
 {
 	struct bch_fs *c = trans->c;
 
 	switch (ret) {
-	case BTREE_INSERT_BTREE_NODE_FULL:
-		ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
-		if (!ret)
-			return 0;
-
-		if (ret == -EINTR)
-			trace_trans_restart_btree_node_split(trans->fn, trace_ip,
-						i->btree_id, &i->path->pos);
+	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_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)
 			break;
 
-		if (bch2_trans_relock(trans))
-			return 0;
-
-		trace_trans_restart_mark_replicas(trans->fn, trace_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 & JOURNAL_WATERMARK_reserved)) {
-			trans->restarted = true;
-			ret = -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)
 			break;
 
-		if (bch2_trans_relock(trans))
-			return 0;
-
-		trace_trans_restart_journal_res_get(trans->fn, trace_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_trans_blocked_journal_reclaim(trans->fn, trace_ip);
+		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)
 			break;
 
-		if (bch2_trans_relock(trans))
-			return 0;
+		ret = bch2_trans_relock(trans);
+		if (ret)
+			trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
+		break;
+	case -BCH_ERR_btree_insert_need_flush_buffer: {
+		struct btree_write_buffer *wb = &c->btree_write_buffer;
+
+		ret = 0;
 
-		trace_trans_restart_journal_reclaim(trans->fn, trace_ip);
-		ret = -EINTR;
+		if (wb->state.nr > wb->size * 3 / 4) {
+			bch2_trans_reset_updates(trans);
+			bch2_trans_unlock(trans);
+
+			mutex_lock(&wb->flush_lock);
+
+			if (wb->state.nr > wb->size * 3 / 4)
+				ret = __bch2_btree_write_buffer_flush(trans,
+						flags|BTREE_INSERT_NOCHECK_RW, true);
+			else
+				mutex_unlock(&wb->flush_lock);
+
+			if (!ret) {
+				trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
+				ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
+			}
+		}
 		break;
+	}
 	default:
 		BUG_ON(ret >= 0);
 		break;
 	}
 
-	BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
-	BUG_ON(ret == -ENOSPC &&
-	       !(trans->flags & BTREE_INSERT_NOWAIT) &&
-	       (trans->flags & BTREE_INSERT_NOFAIL));
+	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)) ||
+	if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
 	    test_bit(BCH_FS_STARTED, &c->flags))
-		return -EROFS;
+		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;
 
-	if (!bch2_trans_relock(trans))
-		return -EINTR;
-
-	percpu_ref_get(&c->writes);
+	bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 	return 0;
 }
 
@@ -1080,18 +1085,20 @@ do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 	return ret;
 }
 
-int __bch2_trans_commit(struct btree_trans *trans)
+int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i = NULL;
+	struct btree_write_buffered_key *wb;
 	unsigned u64s;
 	int ret = 0;
 
 	if (!trans->nr_updates &&
+	    !trans->nr_wb_updates &&
 	    !trans->extra_journal_entries.nr)
 		goto out_reset;
 
-	if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
+	if (flags & BTREE_INSERT_GC_LOCK_HELD)
 		lockdep_assert_held(&c->gc_lock);
 
 	ret = bch2_trans_commit_run_triggers(trans);
@@ -1103,13 +1110,29 @@ int __bch2_trans_commit(struct btree_trans *trans)
 		goto out_reset;
 	}
 
-	if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
-	    unlikely(!percpu_ref_tryget(&c->writes))) {
-		ret = bch2_trans_commit_get_rw_cold(trans);
+	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;
 	}
 
+	if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
+	    mutex_trylock(&c->btree_write_buffer.flush_lock)) {
+		bch2_trans_begin(trans);
+		bch2_trans_unlock(trans);
+
+		ret = __bch2_btree_write_buffer_flush(trans,
+					flags|BTREE_INSERT_NOCHECK_RW, true);
+		if (!ret) {
+			trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
+			ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
+		}
+		goto out;
+	}
+
+	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;
@@ -1118,77 +1141,78 @@ int __bch2_trans_commit(struct btree_trans *trans)
 	trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
 
 	if (trans->journal_transaction_names)
-		trans->journal_u64s += JSET_ENTRY_LOG_U64s;
+		trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
 
 	trans_for_each_update(trans, i) {
-		BUG_ON(!i->path->should_be_locked);
+		EBUG_ON(!i->path->should_be_locked);
 
-		if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
-			trace_trans_restart_upgrade(trans->fn, _RET_IP_,
-						    i->btree_id, &i->path->pos);
-			ret = btree_trans_restart(trans);
+		ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
+		if (unlikely(ret))
 			goto out;
-		}
 
-		BUG_ON(!btree_node_intent_locked(i->path, i->level));
+		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(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
+		    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);
 	}
 
+	trans_for_each_wb_update(trans, wb)
+		trans->journal_u64s += jset_u64s(wb->k.k.u64s);
+
 	if (trans->extra_journal_res) {
 		ret = bch2_disk_reservation_add(c, trans->disk_res,
 				trans->extra_journal_res,
-				(trans->flags & BTREE_INSERT_NOFAIL)
+				(flags & BTREE_INSERT_NOFAIL)
 				? BCH_DISK_RESERVATION_NOFAIL : 0);
 		if (ret)
 			goto err;
 	}
 retry:
-	BUG_ON(trans->restarted);
+	bch2_trans_verify_not_in_restart(trans);
 	memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 
-	ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
+	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(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
-		percpu_ref_put(&c->writes);
+	if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
+		bch2_write_ref_put(c, BCH_WRITE_REF_trans);
 out_reset:
-	trans_for_each_update(trans, i)
-		bch2_path_put(trans, i->path, true);
-
-	trans->extra_journal_res	= 0;
-	trans->nr_updates		= 0;
-	trans->hooks			= NULL;
-	trans->extra_journal_entries.nr	= 0;
-
-	if (trans->fs_usage_deltas) {
-		trans->fs_usage_deltas->used = 0;
-		memset(&trans->fs_usage_deltas->memset_start, 0,
-		       (void *) &trans->fs_usage_deltas->memset_end -
-		       (void *) &trans->fs_usage_deltas->memset_start);
-	}
+	bch2_trans_reset_updates(trans);
 
 	return ret;
 err:
-	ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
+	ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
 	if (ret)
 		goto out;
 
 	goto retry;
 }
 
-static int check_pos_snapshot_overwritten(struct btree_trans *trans,
+static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
 					  enum btree_id id,
 					  struct bpos pos)
 {
@@ -1197,12 +1221,6 @@ static int check_pos_snapshot_overwritten(struct btree_trans *trans,
 	struct bkey_s_c k;
 	int ret;
 
-	if (!btree_type_has_snapshots(id))
-		return 0;
-
-	if (!snapshot_t(c, pos.snapshot)->children[0])
-		return 0;
-
 	bch2_trans_iter_init(trans, &iter, id, pos,
 			     BTREE_ITER_NOT_EXTENTS|
 			     BTREE_ITER_ALL_SNAPSHOTS);
@@ -1215,7 +1233,7 @@ static int check_pos_snapshot_overwritten(struct btree_trans *trans,
 		if (!k.k)
 			break;
 
-		if (bkey_cmp(pos, k.k->p))
+		if (!bkey_eq(pos, k.k->p))
 			break;
 
 		if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
@@ -1228,13 +1246,109 @@ static int check_pos_snapshot_overwritten(struct btree_trans *trans,
 	return ret;
 }
 
+static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
+					  enum btree_id id,
+					  struct bpos pos)
+{
+	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 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;
+
+	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_btree_delete_at(trans, iter, flags);
+	if (ret)
+		return ret;
+
+	*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_bkey_merge(c, bkey_i_to_s(insert), k);
+	return 0;
+}
+
+/*
+ * 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_iter iter;
+	struct bkey_s_c k;
+	u32 snapshot = pos.snapshot;
+	int ret;
+
+	if (!bch2_snapshot_parent(trans->c, pos.snapshot))
+		return 0;
+
+	pos.snapshot++;
+
+	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;
+
+		if (bch2_snapshot_is_ancestor(trans->c, snapshot,
+					      k.k->p.snapshot)) {
+			ret = !bkey_whiteout(k.k);
+			break;
+		}
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
+	return ret;
+}
 int bch2_trans_update_extent(struct btree_trans *trans,
 			     struct btree_iter *orig_iter,
 			     struct bkey_i *insert,
 			     enum btree_update_flags flags)
 {
-	struct bch_fs *c = trans->c;
-	struct btree_iter iter, update_iter;
+	struct btree_iter iter;
 	struct bpos start = bkey_start_pos(&insert->k);
 	struct bkey_i *update;
 	struct bkey_s_c k;
@@ -1251,52 +1365,19 @@ int bch2_trans_update_extent(struct btree_trans *trans,
 	if (!k.k)
 		goto out;
 
-	if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
-		/*
-		 * We can't merge extents if they belong to interior snapshot
-		 * tree nodes, and there's a snapshot in which one extent is
-		 * visible and the other is not - i.e. if visibility is
-		 * different.
-		 *
-		 * Instead of checking if visibilitiy of the two extents is
-		 * different, for now we just check if either has been
-		 * overwritten:
-		 */
-		ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
-		if (ret < 0)
-			goto err;
-		if (ret)
-			goto nomerge1;
-
-		ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
-		if (ret < 0)
-			goto err;
-		if (ret)
-			goto nomerge1;
-
-		update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
-		if ((ret = PTR_ERR_OR_ZERO(update)))
-			goto err;
-
-		bkey_reassemble(update, k);
-
-		if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
-			ret = bch2_btree_delete_at(trans, &iter, flags);
+	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;
-
-			insert = update;
-			goto next;
 		}
-	}
-nomerge1:
-	ret = 0;
-	if (!bkey_cmp(k.k->p, start))
+
 		goto next;
+	}
 
-	while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
-		bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
-		bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
+	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);
 
 		/*
 		 * If we're going to be splitting a compressed extent, note it
@@ -1309,84 +1390,58 @@ nomerge1:
 			trans->extra_journal_res += compressed_sectors;
 
 		if (front_split) {
-			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+			update = bch2_bkey_make_mut(trans, k);
 			if ((ret = PTR_ERR_OR_ZERO(update)))
 				goto err;
 
-			bkey_reassemble(update, k);
-
 			bch2_cut_back(start, 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);
-
+			ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 			if (ret)
 				goto err;
 		}
 
 		if (k.k->p.snapshot != insert->k.p.snapshot &&
 		    (front_split || back_split)) {
-			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+			update = bch2_bkey_make_mut(trans, k);
 			if ((ret = PTR_ERR_OR_ZERO(update)))
 				goto err;
 
-			bkey_reassemble(update, k);
-
 			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);
+			ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 			if (ret)
 				goto err;
 		}
 
-		if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
+		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->k.p.snapshot = insert->k.p.snapshot;
 
-			if (insert->k.p.snapshot != k.k->p.snapshot) {
-				update->k.p.snapshot = insert->k.p.snapshot;
+			if (insert->k.p.snapshot != k.k->p.snapshot ||
+			    (btree_type_has_snapshots(btree_id) &&
+			     need_whiteout_for_snapshot(trans, btree_id, update->k.p)))
 				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);
 
+			ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 			if (ret)
 				goto err;
 		}
 
 		if (back_split) {
-			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+			update = bch2_bkey_make_mut(trans, k);
 			if ((ret = PTR_ERR_OR_ZERO(update)))
 				goto err;
 
-			bkey_reassemble(update, k);
 			bch2_cut_front(insert->k.p, update);
 
 			ret = bch2_trans_update_by_path(trans, iter.path, update,
@@ -1406,22 +1461,10 @@ next:
 	}
 
 	if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
-		ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
-		if (ret < 0)
-			goto err;
+		ret = extent_back_merge(trans, &iter, insert, k);
 		if (ret)
-			goto nomerge2;
-
-		ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
-		if (ret < 0)
 			goto err;
-		if (ret)
-			goto nomerge2;
-
-		bch2_bkey_merge(c, bkey_i_to_s(insert), k);
 	}
-nomerge2:
-	ret = 0;
 out:
 	if (!bkey_deleted(&insert->k)) {
 		/*
@@ -1441,51 +1484,49 @@ err:
 	return ret;
 }
 
-/*
- * 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)
+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);
+
+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_iter iter;
-	struct bkey_s_c k;
-	u32 snapshot = pos.snapshot;
+	struct btree_path *btree_path;
 	int ret;
 
-	if (!bch2_snapshot_parent(trans->c, pos.snapshot))
-		return 0;
-
-	pos.snapshot++;
+	i->key_cache_already_flushed = true;
+	i->flags |= BTREE_TRIGGER_NORUN;
 
-	for_each_btree_key_norestart(trans, iter, btree_id, pos,
-			   BTREE_ITER_ALL_SNAPSHOTS|
-			   BTREE_ITER_NOPRESERVE, k, ret) {
-		if (bkey_cmp(k.k->p, pos))
-			break;
+	btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
+				   BTREE_ITER_INTENT, _THIS_IP_);
 
-		if (bch2_snapshot_is_ancestor(trans->c, snapshot,
-					      k.k->p.snapshot)) {
-			ret = !bkey_whiteout(k.k);
-			break;
-		}
-	}
-	bch2_trans_iter_exit(trans, &iter);
+	ret = bch2_btree_path_traverse(trans, btree_path, 0);
+	if (ret)
+		goto err;
 
+	btree_path_set_should_be_locked(btree_path);
+	ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
+err:
+	bch2_path_put(trans, btree_path, true);
 	return ret;
 }
 
 static 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)
+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 bch_fs *c = trans->c;
 	struct btree_insert_entry *i, n;
+	int cmp;
 
-	BUG_ON(!path->should_be_locked);
-
-	BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
-	BUG_ON(bpos_cmp(k->k.p, path->pos));
+	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,
@@ -1495,7 +1536,7 @@ bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
 		.cached		= path->cached,
 		.path		= path,
 		.k		= k,
-		.ip_allocated	= _RET_IP_,
+		.ip_allocated	= ip,
 	};
 
 #ifdef CONFIG_BCACHEFS_DEBUG
@@ -1508,13 +1549,14 @@ bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
 	 * Pending updates are kept sorted: first, find position of new update,
 	 * then delete/trim any updates the new update overwrites:
 	 */
-	trans_for_each_update(trans, i)
-		if (btree_insert_entry_cmp(&n, i) <= 0)
+	trans_for_each_update(trans, i) {
+		cmp = btree_insert_entry_cmp(&n, i);
+		if (cmp <= 0)
 			break;
+	}
 
-	if (i < trans->updates + trans->nr_updates &&
-	    !btree_insert_entry_cmp(&n, i)) {
-		BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
+	if (!cmp && i < trans->updates + trans->nr_updates) {
+		EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
 
 		bch2_path_put(trans, i->path, true);
 		i->flags	= n.flags;
@@ -1526,24 +1568,44 @@ bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
 		array_insert_item(trans->updates, trans->nr_updates,
 				  i - trans->updates, n);
 
-		i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
+		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(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
+		if (unlikely(trans->journal_replay_not_finished)) {
 			struct bkey_i *j_k =
-				bch2_journal_keys_peek(c, n.btree_id, n.level, k->k.p);
+				bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
 
-			if (j_k && !bpos_cmp(j_k->k.p, i->k->k.p)) {
+			if (j_k) {
 				i->old_k = j_k->k;
 				i->old_v = &j_k->v;
 			}
 		}
 	}
 
-	__btree_path_get(n.path, true);
+	__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)
 {
@@ -1565,13 +1627,16 @@ int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter
 			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_cmp(iter->key_cache_path->pos, k->k.p)) {
+		    !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,
@@ -1591,12 +1656,11 @@ int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter
 			ck = (void *) iter->key_cache_path->l[0].b;
 
 			if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
-				trace_trans_restart_key_cache_raced(trans->fn, _RET_IP_);
-				btree_trans_restart(trans);
-				return -EINTR;
+				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);
 			}
 
-			iter->key_cache_path->should_be_locked = true;
+			btree_path_set_should_be_locked(iter->key_cache_path);
 		}
 
 		path = iter->key_cache_path;
@@ -1605,6 +1669,59 @@ int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter
 	return bch2_trans_update_by_path(trans, path, k, flags);
 }
 
+int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
+					    enum btree_id btree,
+					    struct bkey_i *k)
+{
+	struct btree_write_buffered_key *i;
+	int ret;
+
+	EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
+	EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
+
+	trans_for_each_wb_update(trans, i) {
+		if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
+			bkey_copy(&i->k, k);
+			return 0;
+		}
+	}
+
+	if (!trans->wb_updates ||
+	    trans->nr_wb_updates == trans->wb_updates_size) {
+		struct btree_write_buffered_key *u;
+
+		if (trans->nr_wb_updates == trans->wb_updates_size) {
+			struct btree_transaction_stats *s = btree_trans_stats(trans);
+
+			BUG_ON(trans->wb_updates_size > U8_MAX / 2);
+			trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
+			if (s)
+				s->wb_updates_size = trans->wb_updates_size;
+		}
+
+		u = bch2_trans_kmalloc_nomemzero(trans,
+					trans->wb_updates_size *
+					sizeof(struct btree_write_buffered_key));
+		ret = PTR_ERR_OR_ZERO(u);
+		if (ret)
+			return ret;
+
+		if (trans->nr_wb_updates)
+			memcpy(u, trans->wb_updates, trans->nr_wb_updates *
+			       sizeof(struct btree_write_buffered_key));
+		trans->wb_updates = u;
+	}
+
+	trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
+		.btree	= btree,
+	};
+
+	bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
+	trans->nr_wb_updates++;
+
+	return 0;
+}
+
 void bch2_trans_commit_hook(struct btree_trans *trans,
 			    struct btree_trans_commit_hook *h)
 {
@@ -1612,16 +1729,33 @@ void bch2_trans_commit_hook(struct btree_trans *trans,
 	trans->hooks = h;
 }
 
-int __bch2_btree_insert(struct btree_trans *trans,
-			enum btree_id id, struct bkey_i *k)
+int bch2_btree_insert_nonextent(struct btree_trans *trans,
+				enum btree_id btree, struct bkey_i *k,
+				enum btree_update_flags flags)
+{
+	struct btree_iter iter;
+	int ret;
+
+	bch2_trans_iter_init(trans, &iter, btree, k->k.p,
+			     BTREE_ITER_NOT_EXTENTS|
+			     BTREE_ITER_INTENT);
+	ret   = bch2_btree_iter_traverse(&iter) ?:
+		bch2_trans_update(trans, &iter, k, flags);
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
+			struct bkey_i *k, enum btree_update_flags flags)
 {
 	struct btree_iter iter;
 	int ret;
 
 	bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
+			     BTREE_ITER_CACHED|
 			     BTREE_ITER_INTENT);
 	ret   = bch2_btree_iter_traverse(&iter) ?:
-		bch2_trans_update(trans, &iter, k, 0);
+		bch2_trans_update(trans, &iter, k, flags);
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
@@ -1639,11 +1773,11 @@ int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
 		      u64 *journal_seq, int flags)
 {
 	return bch2_trans_do(c, disk_res, journal_seq, flags,
-			     __bch2_btree_insert(&trans, id, k));
+			     __bch2_btree_insert(&trans, id, k, 0));
 }
 
-int bch2_btree_delete_at(struct btree_trans *trans,
-			 struct btree_iter *iter, unsigned update_flags)
+int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
+				unsigned len, unsigned update_flags)
 {
 	struct bkey_i *k;
 
@@ -1653,28 +1787,36 @@ int bch2_btree_delete_at(struct btree_trans *trans,
 
 	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)
 {
+	u32 restart_count = trans->restart_count;
 	struct btree_iter iter;
 	struct bkey_s_c k;
 	int ret = 0;
 
 	bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
-retry:
-	while ((bch2_trans_begin(trans),
-	       (k = bch2_btree_iter_peek(&iter)).k) &&
-	       !(ret = bkey_err(k)) &&
-	       bkey_cmp(iter.pos, end) < 0) {
+	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;
 
+		ret = bkey_err(k);
+		if (ret)
+			goto err;
+
 		bkey_init(&delete.k);
 
 		/*
@@ -1693,34 +1835,33 @@ retry:
 		 */
 		delete.k.p = iter.pos;
 
-		if (iter.flags & BTREE_ITER_IS_EXTENTS) {
-			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);
+		if (iter.flags & BTREE_ITER_IS_EXTENTS)
+			bch2_key_resize(&delete.k,
+					bpos_min(end, k.k->p).offset -
+					iter.pos.offset);
 
-			ret = bch2_extent_trim_atomic(trans, &iter, &delete);
-			if (ret)
-				break;
-		}
-
-		ret   = bch2_trans_update(trans, &iter, &delete, 0) ?:
+		ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
 			bch2_trans_commit(trans, &disk_res, journal_seq,
-					  BTREE_INSERT_NOFAIL|
-					  update_flags);
+					  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);
+
+		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+			ret = 0;
 		if (ret)
 			break;
 	}
-
-	if (ret == -EINTR) {
-		ret = 0;
-		goto retry;
-	}
-
 	bch2_trans_iter_exit(trans, &iter);
+
+	if (!ret && trans_was_restarted(trans, restart_count))
+		ret = -BCH_ERR_transaction_restart_nested;
 	return ret;
 }
 
@@ -1734,23 +1875,33 @@ int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 			    unsigned update_flags,
 			    u64 *journal_seq)
 {
-	return bch2_trans_do(c, NULL, journal_seq, 0,
-			     bch2_btree_delete_range_trans(&trans, id, start, end,
-							   update_flags, 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;
 }
 
-int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
+static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
 {
-	unsigned len = strlen(msg);
-	unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
+	struct printbuf buf = PRINTBUF;
 	struct jset_entry_log *l;
+	unsigned u64s;
 	int ret;
 
-	ret = darray_make_room(trans->extra_journal_entries, jset_u64s(u64s));
+	prt_vprintf(&buf, fmt, args);
+	ret = buf.allocation_failure ? -ENOMEM : 0;
 	if (ret)
-		return ret;
+		goto err;
+
+	u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
 
-	l = (void *) &darray_top(trans->extra_journal_entries);
+	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;
@@ -1758,10 +1909,44 @@ int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
 	l->entry.pad[0]		= 0;
 	l->entry.pad[1]		= 0;
 	l->entry.pad[2]		= 0;
-	memcpy(l->d, msg, len);
-	while (len & 7)
-		l->d[len++] = '\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;
 
-	trans->extra_journal_entries.nr += jset_u64s(u64s);
-	return 0;
 }