X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_update_interior.c;h=c741150e68af971d144e752df543078bf1396bf8;hb=8d5e53b88aaafe7c01fc369e52dbd1fc8955a77d;hp=5ce91ae6a69e800f9a839701c1ad7fc98c83f966;hpb=ae43a58d97fc00e31770142da832fb8a249808eb;p=bcachefs-tools-debian

diff --git a/libbcachefs/btree_update_interior.c b/libbcachefs/btree_update_interior.c
index 5ce91ae..c741150 100644
--- a/libbcachefs/btree_update_interior.c
+++ b/libbcachefs/btree_update_interior.c
@@ -5,23 +5,24 @@
 #include "bkey_methods.h"
 #include "btree_cache.h"
 #include "btree_gc.h"
+#include "btree_journal_iter.h"
 #include "btree_update.h"
 #include "btree_update_interior.h"
 #include "btree_io.h"
 #include "btree_iter.h"
 #include "btree_locking.h"
 #include "buckets.h"
+#include "clock.h"
 #include "error.h"
 #include "extents.h"
 #include "journal.h"
 #include "journal_reclaim.h"
 #include "keylist.h"
-#include "recovery.h"
 #include "replicas.h"
 #include "super-io.h"
+#include "trace.h"
 
 #include <linux/random.h>
-#include <trace/events/bcachefs.h>
 
 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
 				  struct btree_path *, struct btree *,
@@ -72,7 +73,7 @@ static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
 			break;
 		bp = bkey_s_c_to_btree_ptr_v2(k);
 
-		if (bpos_cmp(next_node, bp.v->min_key)) {
+		if (!bpos_eq(next_node, bp.v->min_key)) {
 			bch2_dump_btree_node(c, b);
 			bch2_bpos_to_text(&buf1, next_node);
 			bch2_bpos_to_text(&buf2, bp.v->min_key);
@@ -82,7 +83,7 @@ static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
 		bch2_btree_node_iter_advance(&iter, b);
 
 		if (bch2_btree_node_iter_end(&iter)) {
-			if (bpos_cmp(k.k->p, b->key.k.p)) {
+			if (!bpos_eq(k.k->p, b->key.k.p)) {
 				bch2_dump_btree_node(c, b);
 				bch2_bpos_to_text(&buf1, b->key.k.p);
 				bch2_bpos_to_text(&buf2, k.k->p);
@@ -161,6 +162,7 @@ static void __btree_node_free(struct bch_fs *c, struct btree *b)
 {
 	trace_and_count(c, btree_node_free, c, b);
 
+	BUG_ON(btree_node_write_blocked(b));
 	BUG_ON(btree_node_dirty(b));
 	BUG_ON(btree_node_need_write(b));
 	BUG_ON(b == btree_node_root(c, b));
@@ -186,7 +188,7 @@ static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 	bch2_btree_node_hash_remove(&c->btree_cache, b);
 	__btree_node_free(c, b);
 	six_unlock_write(&b->c.lock);
-	mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
+	mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 
 	trans_for_each_path(trans, path)
 		if (path->l[level].b == b) {
@@ -241,19 +243,14 @@ static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 	struct bch_fs *c = trans->c;
 	struct write_point *wp;
 	struct btree *b;
-	__BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
+	BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 	struct open_buckets ob = { .nr = 0 };
 	struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
-	unsigned nr_reserve;
-	enum alloc_reserve alloc_reserve;
-
-	if (flags & BTREE_INSERT_USE_RESERVE) {
-		nr_reserve	= 0;
-		alloc_reserve	= RESERVE_btree_movinggc;
-	} else {
-		nr_reserve	= BTREE_NODE_RESERVE;
-		alloc_reserve	= RESERVE_btree;
-	}
+	enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
+	unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
+		? BTREE_NODE_RESERVE
+		: 0;
+	int ret;
 
 	mutex_lock(&c->btree_reserve_cache_lock);
 	if (c->btree_reserve_cache_nr > nr_reserve) {
@@ -268,7 +265,7 @@ static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 	mutex_unlock(&c->btree_reserve_cache_lock);
 
 retry:
-	wp = bch2_alloc_sectors_start_trans(trans,
+	ret = bch2_alloc_sectors_start_trans(trans,
 				      c->opts.metadata_target ?:
 				      c->opts.foreground_target,
 				      0,
@@ -276,9 +273,9 @@ retry:
 				      &devs_have,
 				      res->nr_replicas,
 				      c->opts.metadata_replicas_required,
-				      alloc_reserve, 0, cl);
-	if (IS_ERR(wp))
-		return ERR_CAST(wp);
+				      watermark, 0, cl, &wp);
+	if (unlikely(ret))
+		return ERR_PTR(ret);
 
 	if (wp->sectors_free < btree_sectors(c)) {
 		struct open_bucket *ob;
@@ -298,7 +295,7 @@ retry:
 	bch2_open_bucket_get(c, wp, &ob);
 	bch2_alloc_sectors_done(c, wp);
 mem_alloc:
-	b = bch2_btree_node_mem_alloc(c, interior_node);
+	b = bch2_btree_node_mem_alloc(trans, interior_node);
 	six_unlock_write(&b->c.lock);
 	six_unlock_intent(&b->c.lock);
 
@@ -361,6 +358,7 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 	BUG_ON(ret);
 
 	trace_and_count(c, btree_node_alloc, c, b);
+	bch2_increment_clock(c, btree_sectors(c), WRITE);
 	return b;
 }
 
@@ -377,14 +375,19 @@ static void btree_set_max(struct btree *b, struct bpos pos)
 	b->data->max_key = pos;
 }
 
-struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
-						  struct btree_trans *trans,
-						  struct btree *b,
-						  struct bkey_format format)
+static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
+						       struct btree_trans *trans,
+						       struct btree *b)
 {
-	struct btree *n;
+	struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
+	struct bkey_format format = bch2_btree_calc_format(b);
 
-	n = bch2_btree_node_alloc(as, trans, b->c.level);
+	/*
+	 * The keys might expand with the new format - if they wouldn't fit in
+	 * the btree node anymore, use the old format for now:
+	 */
+	if (!bch2_btree_node_format_fits(as->c, b, &format))
+		format = b->format;
 
 	SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 
@@ -397,27 +400,9 @@ struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 	bch2_btree_sort_into(as->c, n, b);
 
 	btree_node_reset_sib_u64s(n);
-
-	n->key.k.p = b->key.k.p;
 	return n;
 }
 
-static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
-						       struct btree_trans *trans,
-						       struct btree *b)
-{
-	struct bkey_format new_f = bch2_btree_calc_format(b);
-
-	/*
-	 * The keys might expand with the new format - if they wouldn't fit in
-	 * the btree node anymore, use the old format for now:
-	 */
-	if (!bch2_btree_node_format_fits(as->c, b, &new_f))
-		new_f = b->format;
-
-	return __bch2_btree_node_alloc_replacement(as, trans, b, new_f);
-}
-
 static struct btree *__btree_root_alloc(struct btree_update *as,
 				struct btree_trans *trans, unsigned level)
 {
@@ -656,11 +641,10 @@ static void btree_update_nodes_written(struct btree_update *as)
 	 * which may require allocations as well.
 	 */
 	ret = commit_do(&trans, &as->disk_res, &journal_seq,
+			BCH_WATERMARK_reclaim|
 			BTREE_INSERT_NOFAIL|
 			BTREE_INSERT_NOCHECK_RW|
-			BTREE_INSERT_USE_RESERVE|
-			BTREE_INSERT_JOURNAL_RECLAIM|
-			JOURNAL_WATERMARK_reserved,
+			BTREE_INSERT_JOURNAL_RECLAIM,
 			btree_update_nodes_written_trans(&trans, as));
 	bch2_trans_unlock(&trans);
 
@@ -697,7 +681,8 @@ err:
 		bch2_trans_unlock(&trans);
 		btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_intent);
 		mark_btree_node_locked(&trans, path, b->c.level, SIX_LOCK_intent);
-		bch2_btree_path_level_init(&trans, path, b);
+		path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
+		path->l[b->c.level].b = b;
 
 		bch2_btree_node_lock_write_nofail(&trans, path, &b->c);
 
@@ -735,7 +720,7 @@ err:
 
 		mutex_unlock(&c->btree_interior_update_lock);
 
-		mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
+		mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 		six_unlock_write(&b->c.lock);
 
 		btree_node_write_if_need(c, b, SIX_LOCK_intent);
@@ -818,6 +803,7 @@ static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 
 	BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 	BUG_ON(!btree_node_dirty(b));
+	BUG_ON(!b->c.level);
 
 	as->mode	= BTREE_INTERIOR_UPDATING_NODE;
 	as->b		= b;
@@ -987,6 +973,7 @@ static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 
 	clear_btree_node_dirty_acct(c, b);
 	clear_btree_node_need_write(b);
+	clear_btree_node_write_blocked(b);
 
 	/*
 	 * Does this node have unwritten data that has a pin on the journal?
@@ -1055,14 +1042,24 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 		? BCH_DISK_RESERVATION_NOFAIL : 0;
 	unsigned nr_nodes[2] = { 0, 0 };
 	unsigned update_level = level;
-	int journal_flags = flags & JOURNAL_WATERMARK_MASK;
+	enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
+	unsigned journal_flags = 0;
 	int ret = 0;
 	u32 restart_count = trans->restart_count;
 
 	BUG_ON(!path->should_be_locked);
 
+	if (watermark == BCH_WATERMARK_copygc)
+		watermark = BCH_WATERMARK_btree_copygc;
+	if (watermark < BCH_WATERMARK_btree)
+		watermark = BCH_WATERMARK_btree;
+
+	flags &= ~BCH_WATERMARK_MASK;
+	flags |= watermark;
+
 	if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
 		journal_flags |= JOURNAL_RES_GET_NONBLOCK;
+	journal_flags |= watermark;
 
 	while (1) {
 		nr_nodes[!!update_level] += 1 + split;
@@ -1083,22 +1080,20 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 					BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
 			break;
 
-		split = true;
+		split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
 	}
 
 	if (flags & BTREE_INSERT_GC_LOCK_HELD)
 		lockdep_assert_held(&c->gc_lock);
 	else if (!down_read_trylock(&c->gc_lock)) {
-		bch2_trans_unlock(trans);
-		down_read(&c->gc_lock);
-		ret = bch2_trans_relock(trans);
+		ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
 		if (ret) {
 			up_read(&c->gc_lock);
 			return ERR_PTR(ret);
 		}
 	}
 
-	as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
+	as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
 	memset(as, 0, sizeof(*as));
 	closure_init(&as->cl, NULL);
 	as->c		= c;
@@ -1134,23 +1129,19 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 				      BTREE_UPDATE_JOURNAL_RES,
 				      journal_flags|JOURNAL_RES_GET_NONBLOCK);
 	if (ret) {
-		bch2_trans_unlock(trans);
-
 		if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
 			ret = -BCH_ERR_journal_reclaim_would_deadlock;
 			goto err;
 		}
 
-		ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
+		ret = drop_locks_do(trans,
+			bch2_journal_preres_get(&c->journal, &as->journal_preres,
 					      BTREE_UPDATE_JOURNAL_RES,
-					      journal_flags);
-		if (ret) {
+					      journal_flags));
+		if (ret == -BCH_ERR_journal_preres_get_blocked) {
 			trace_and_count(c, trans_restart_journal_preres_get, trans, _RET_IP_, journal_flags);
 			ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get);
-			goto err;
 		}
-
-		ret = bch2_trans_relock(trans);
 		if (ret)
 			goto err;
 	}
@@ -1167,6 +1158,17 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 	    bch2_err_matches(ret, ENOMEM)) {
 		struct closure cl;
 
+		/*
+		 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
+		 * flag
+		 */
+		if (bch2_err_matches(ret, ENOSPC) &&
+		    (flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
+		    watermark != BCH_WATERMARK_reclaim) {
+			ret = -BCH_ERR_journal_reclaim_would_deadlock;
+			goto err;
+		}
+
 		closure_init_stack(&cl);
 
 		do {
@@ -1174,11 +1176,12 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 
 			bch2_trans_unlock(trans);
 			closure_sync(&cl);
-		} while (ret == -EAGAIN);
+		} while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
 	}
 
 	if (ret) {
-		trace_and_count(c, btree_reserve_get_fail, trans->fn, _RET_IP_, nr_nodes[0] + nr_nodes[1]);
+		trace_and_count(c, btree_reserve_get_fail, trans->fn,
+				_RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
 		goto err;
 	}
 
@@ -1207,7 +1210,7 @@ static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
 	       (b->c.level < btree_node_root(c, b)->c.level ||
 		!btree_node_dying(btree_node_root(c, b))));
 
-	btree_node_root(c, b) = b;
+	bch2_btree_id_root(c, b->c.btree_id)->b = b;
 	mutex_unlock(&c->btree_root_lock);
 
 	bch2_recalc_btree_reserve(c);
@@ -1269,6 +1272,7 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
 	struct bch_fs *c = as->c;
 	struct bkey_packed *k;
 	struct printbuf buf = PRINTBUF;
+	unsigned long old, new, v;
 
 	BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
 	       !btree_ptr_sectors_written(insert));
@@ -1306,7 +1310,15 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
 
 	bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
 	set_btree_node_dirty_acct(c, b);
-	set_btree_node_need_write(b);
+
+	v = READ_ONCE(b->flags);
+	do {
+		old = new = v;
+
+		new &= ~BTREE_WRITE_TYPE_MASK;
+		new |= BTREE_WRITE_interior;
+		new |= 1 << BTREE_NODE_need_write;
+	} while ((v = cmpxchg(&b->flags, old, new)) != old);
 
 	printbuf_exit(&buf);
 }
@@ -1329,8 +1341,12 @@ __bch2_btree_insert_keys_interior(struct btree_update *as,
 		;
 
 	while (!bch2_keylist_empty(keys)) {
-		bch2_insert_fixup_btree_ptr(as, trans, path, b,
-				&node_iter, bch2_keylist_front(keys));
+		struct bkey_i *k = bch2_keylist_front(keys);
+
+		if (bpos_gt(k->k.p, b->key.k.p))
+			break;
+
+		bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, k);
 		bch2_keylist_pop_front(keys);
 	}
 }
@@ -1339,109 +1355,91 @@ __bch2_btree_insert_keys_interior(struct btree_update *as,
  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
  * node)
  */
-static struct btree *__btree_split_node(struct btree_update *as,
-					struct btree_trans *trans,
-					struct btree *n1)
+static void __btree_split_node(struct btree_update *as,
+			       struct btree_trans *trans,
+			       struct btree *b,
+			       struct btree *n[2])
 {
-	struct bkey_format_state s;
-	size_t nr_packed = 0, nr_unpacked = 0;
-	struct btree *n2;
-	struct bset *set1, *set2;
-	struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
-	struct bpos n1_pos;
+	struct bkey_packed *k;
+	struct bpos n1_pos = POS_MIN;
+	struct btree_node_iter iter;
+	struct bset *bsets[2];
+	struct bkey_format_state format[2];
+	struct bkey_packed *out[2];
+	struct bkey uk;
+	unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
+	int i;
 
-	n2 = bch2_btree_node_alloc(as, trans, n1->c.level);
+	for (i = 0; i < 2; i++) {
+		BUG_ON(n[i]->nsets != 1);
 
-	n2->data->max_key	= n1->data->max_key;
-	n2->data->format	= n1->format;
-	SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
-	n2->key.k.p = n1->key.k.p;
+		bsets[i] = btree_bset_first(n[i]);
+		out[i] = bsets[i]->start;
 
-	set1 = btree_bset_first(n1);
-	set2 = btree_bset_first(n2);
+		SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
+		bch2_bkey_format_init(&format[i]);
+	}
 
-	/*
-	 * Has to be a linear search because we don't have an auxiliary
-	 * search tree yet
-	 */
-	k = set1->start;
-	while (1) {
-		struct bkey_packed *n = bkey_next(k);
+	u64s = 0;
+	for_each_btree_node_key(b, k, &iter) {
+		if (bkey_deleted(k))
+			continue;
+
+		i = u64s >= n1_u64s;
+		u64s += k->u64s;
+		uk = bkey_unpack_key(b, k);
+		if (!i)
+			n1_pos = uk.p;
+		bch2_bkey_format_add_key(&format[i], &uk);
+	}
 
-		if (n == vstruct_last(set1))
-			break;
-		if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
-			break;
+	btree_set_min(n[0], b->data->min_key);
+	btree_set_max(n[0], n1_pos);
+	btree_set_min(n[1], bpos_successor(n1_pos));
+	btree_set_max(n[1], b->data->max_key);
 
-		if (bkey_packed(k))
-			nr_packed++;
-		else
-			nr_unpacked++;
+	for (i = 0; i < 2; i++) {
+		bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
+		bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
 
-		prev = k;
-		k = n;
+		n[i]->data->format = bch2_bkey_format_done(&format[i]);
+		btree_node_set_format(n[i], n[i]->data->format);
 	}
 
-	BUG_ON(!prev);
-	set2_start	= k;
-	set2_end	= vstruct_last(set1);
-
-	set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
-	set_btree_bset_end(n1, n1->set);
+	u64s = 0;
+	for_each_btree_node_key(b, k, &iter) {
+		if (bkey_deleted(k))
+			continue;
 
-	n1->nr.live_u64s	= le16_to_cpu(set1->u64s);
-	n1->nr.bset_u64s[0]	= le16_to_cpu(set1->u64s);
-	n1->nr.packed_keys	= nr_packed;
-	n1->nr.unpacked_keys	= nr_unpacked;
+		i = u64s >= n1_u64s;
+		u64s += k->u64s;
 
-	n1_pos = bkey_unpack_pos(n1, prev);
-	if (as->c->sb.version < bcachefs_metadata_version_snapshot)
-		n1_pos.snapshot = U32_MAX;
+		if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
+					? &b->format: &bch2_bkey_format_current, k))
+			out[i]->format = KEY_FORMAT_LOCAL_BTREE;
+		else
+			bch2_bkey_unpack(b, (void *) out[i], k);
 
-	btree_set_max(n1, n1_pos);
-	btree_set_min(n2, bpos_successor(n1->key.k.p));
+		out[i]->needs_whiteout = false;
 
-	bch2_bkey_format_init(&s);
-	bch2_bkey_format_add_pos(&s, n2->data->min_key);
-	bch2_bkey_format_add_pos(&s, n2->data->max_key);
-
-	for (k = set2_start; k != set2_end; k = bkey_next(k)) {
-		struct bkey uk = bkey_unpack_key(n1, k);
-		bch2_bkey_format_add_key(&s, &uk);
+		btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
+		out[i] = bkey_p_next(out[i]);
 	}
 
-	n2->data->format = bch2_bkey_format_done(&s);
-	btree_node_set_format(n2, n2->data->format);
-
-	out = set2->start;
-	memset(&n2->nr, 0, sizeof(n2->nr));
-
-	for (k = set2_start; k != set2_end; k = bkey_next(k)) {
-		BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
-				       ? &n1->format : &bch2_bkey_format_current, k));
-		out->format = KEY_FORMAT_LOCAL_BTREE;
-		btree_keys_account_key_add(&n2->nr, 0, out);
-		out = bkey_next(out);
-	}
+	for (i = 0; i < 2; i++) {
+		bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
 
-	set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
-	set_btree_bset_end(n2, n2->set);
+		BUG_ON(!bsets[i]->u64s);
 
-	BUG_ON(!set1->u64s);
-	BUG_ON(!set2->u64s);
+		set_btree_bset_end(n[i], n[i]->set);
 
-	btree_node_reset_sib_u64s(n1);
-	btree_node_reset_sib_u64s(n2);
+		btree_node_reset_sib_u64s(n[i]);
 
-	bch2_verify_btree_nr_keys(n1);
-	bch2_verify_btree_nr_keys(n2);
+		bch2_verify_btree_nr_keys(n[i]);
 
-	if (n1->c.level) {
-		btree_node_interior_verify(as->c, n1);
-		btree_node_interior_verify(as->c, n2);
+		if (b->c.level)
+			btree_node_interior_verify(as->c, n[i]);
 	}
-
-	return n2;
 }
 
 /*
@@ -1461,41 +1459,16 @@ static void btree_split_insert_keys(struct btree_update *as,
 				    struct btree *b,
 				    struct keylist *keys)
 {
-	struct btree_node_iter node_iter;
-	struct bkey_i *k = bch2_keylist_front(keys);
-	struct bkey_packed *src, *dst, *n;
-	struct bset *i;
+	if (!bch2_keylist_empty(keys) &&
+	    bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
+		struct btree_node_iter node_iter;
 
-	bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
+		bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
 
-	__bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
+		__bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
 
-	/*
-	 * We can't tolerate whiteouts here - with whiteouts there can be
-	 * duplicate keys, and it would be rather bad if we picked a duplicate
-	 * for the pivot:
-	 */
-	i = btree_bset_first(b);
-	src = dst = i->start;
-	while (src != vstruct_last(i)) {
-		n = bkey_next(src);
-		if (!bkey_deleted(src)) {
-			memmove_u64s_down(dst, src, src->u64s);
-			dst = bkey_next(dst);
-		}
-		src = n;
+		btree_node_interior_verify(as->c, b);
 	}
-
-	/* Also clear out the unwritten whiteouts area: */
-	b->whiteout_u64s = 0;
-
-	i->u64s = cpu_to_le16((u64 *) dst - i->_data);
-	set_btree_bset_end(b, b->set);
-
-	BUG_ON(b->nsets != 1 ||
-	       b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
-
-	btree_node_interior_verify(as->c, b);
 }
 
 static int btree_split(struct btree_update *as, struct btree_trans *trans,
@@ -1514,15 +1487,21 @@ static int btree_split(struct btree_update *as, struct btree_trans *trans,
 
 	bch2_btree_interior_update_will_free_node(as, b);
 
-	n1 = bch2_btree_node_alloc_replacement(as, trans, b);
-
-	if (keys)
-		btree_split_insert_keys(as, trans, path, n1, keys);
+	if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
+		struct btree *n[2];
 
-	if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
 		trace_and_count(c, btree_node_split, c, b);
 
-		n2 = __btree_split_node(as, trans, n1);
+		n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
+		n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
+
+		__btree_split_node(as, trans, b, n);
+
+		if (keys) {
+			btree_split_insert_keys(as, trans, path, n1, keys);
+			btree_split_insert_keys(as, trans, path, n2, keys);
+			BUG_ON(!bch2_keylist_empty(keys));
+		}
 
 		bch2_btree_build_aux_trees(n2);
 		bch2_btree_build_aux_trees(n1);
@@ -1571,6 +1550,13 @@ static int btree_split(struct btree_update *as, struct btree_trans *trans,
 	} else {
 		trace_and_count(c, btree_node_compact, c, b);
 
+		n1 = bch2_btree_node_alloc_replacement(as, trans, b);
+
+		if (keys) {
+			btree_split_insert_keys(as, trans, path, n1, keys);
+			BUG_ON(!bch2_keylist_empty(keys));
+		}
+
 		bch2_btree_build_aux_trees(n1);
 		bch2_btree_update_add_new_node(as, n1);
 		six_unlock_write(&n1->c.lock);
@@ -1702,9 +1688,6 @@ static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *t
 	BUG_ON(!as || as->b);
 	bch2_verify_keylist_sorted(keys);
 
-	if (!(local_clock() & 63))
-		return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
-
 	ret = bch2_btree_node_lock_write(trans, path, &b->c);
 	if (ret)
 		return ret;
@@ -1741,8 +1724,10 @@ split:
 	 * We could attempt to avoid the transaction restart, by calling
 	 * bch2_btree_path_upgrade() and allocating more nodes:
 	 */
-	if (b->c.level >= as->update_level)
+	if (b->c.level >= as->update_level) {
+		trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
 		return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
+	}
 
 	return btree_split(as, trans, path, b, keys, flags);
 }
@@ -1798,8 +1783,8 @@ int __bch2_foreground_maybe_merge(struct btree_trans *trans,
 
 	b = path->l[level].b;
 
-	if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
-	    (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
+	if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
+	    (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
 		b->sib_u64s[sib] = U16_MAX;
 		return 0;
 	}
@@ -1832,7 +1817,7 @@ int __bch2_foreground_maybe_merge(struct btree_trans *trans,
 		next = m;
 	}
 
-	if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
+	if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
 		struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 
 		bch2_bpos_to_text(&buf1, prev->data->max_key);
@@ -1874,9 +1859,7 @@ int __bch2_foreground_maybe_merge(struct btree_trans *trans,
 
 	parent = btree_node_parent(path, b);
 	as = bch2_btree_update_start(trans, path, level, false,
-			 BTREE_INSERT_NOFAIL|
-			 BTREE_INSERT_USE_RESERVE|
-			 flags);
+				     BTREE_INSERT_NOFAIL|flags);
 	ret = PTR_ERR_OR_ZERO(as);
 	if (ret)
 		goto err;
@@ -2022,6 +2005,7 @@ err:
 struct async_btree_rewrite {
 	struct bch_fs		*c;
 	struct work_struct	work;
+	struct list_head	list;
 	enum btree_id		btree_id;
 	unsigned		level;
 	struct bpos		pos;
@@ -2031,6 +2015,7 @@ struct async_btree_rewrite {
 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
 					  struct async_btree_rewrite *a)
 {
+	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
 	struct btree *b;
 	int ret;
@@ -2042,8 +2027,18 @@ static int async_btree_node_rewrite_trans(struct btree_trans *trans,
 	if (ret)
 		goto out;
 
-	if (!b || b->data->keys.seq != a->seq)
+	if (!b || b->data->keys.seq != a->seq) {
+		struct printbuf buf = PRINTBUF;
+
+		if (b)
+			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
+		else
+			prt_str(&buf, "(null");
+		bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
+			 __func__, a->seq, buf.buf);
+		printbuf_exit(&buf);
 		goto out;
+	}
 
 	ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 out:
@@ -2052,28 +2047,29 @@ out:
 	return ret;
 }
 
-void async_btree_node_rewrite_work(struct work_struct *work)
+static void async_btree_node_rewrite_work(struct work_struct *work)
 {
 	struct async_btree_rewrite *a =
 		container_of(work, struct async_btree_rewrite, work);
 	struct bch_fs *c = a->c;
+	int ret;
 
-	bch2_trans_do(c, NULL, NULL, 0,
+	ret = bch2_trans_do(c, NULL, NULL, 0,
 		      async_btree_node_rewrite_trans(&trans, a));
-	percpu_ref_put(&c->writes);
+	if (ret)
+		bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
+	bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
 	kfree(a);
 }
 
 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
 {
 	struct async_btree_rewrite *a;
-
-	if (!percpu_ref_tryget_live(&c->writes))
-		return;
+	int ret;
 
 	a = kmalloc(sizeof(*a), GFP_NOFS);
 	if (!a) {
-		percpu_ref_put(&c->writes);
+		bch_err(c, "%s: error allocating memory", __func__);
 		return;
 	}
 
@@ -2082,15 +2078,68 @@ void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
 	a->level	= b->c.level;
 	a->pos		= b->key.k.p;
 	a->seq		= b->data->keys.seq;
-
 	INIT_WORK(&a->work, async_btree_node_rewrite_work);
+
+	if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
+		mutex_lock(&c->pending_node_rewrites_lock);
+		list_add(&a->list, &c->pending_node_rewrites);
+		mutex_unlock(&c->pending_node_rewrites_lock);
+		return;
+	}
+
+	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
+		if (test_bit(BCH_FS_STARTED, &c->flags)) {
+			bch_err(c, "%s: error getting c->writes ref", __func__);
+			kfree(a);
+			return;
+		}
+
+		ret = bch2_fs_read_write_early(c);
+		if (ret) {
+			bch_err(c, "%s: error going read-write: %s",
+				__func__, bch2_err_str(ret));
+			kfree(a);
+			return;
+		}
+
+		bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
+	}
+
 	queue_work(c->btree_interior_update_worker, &a->work);
 }
 
+void bch2_do_pending_node_rewrites(struct bch_fs *c)
+{
+	struct async_btree_rewrite *a, *n;
+
+	mutex_lock(&c->pending_node_rewrites_lock);
+	list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
+		list_del(&a->list);
+
+		bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
+		queue_work(c->btree_interior_update_worker, &a->work);
+	}
+	mutex_unlock(&c->pending_node_rewrites_lock);
+}
+
+void bch2_free_pending_node_rewrites(struct bch_fs *c)
+{
+	struct async_btree_rewrite *a, *n;
+
+	mutex_lock(&c->pending_node_rewrites_lock);
+	list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
+		list_del(&a->list);
+
+		kfree(a);
+	}
+	mutex_unlock(&c->pending_node_rewrites_lock);
+}
+
 static int __bch2_btree_node_update_key(struct btree_trans *trans,
 					struct btree_iter *iter,
 					struct btree *b, struct btree *new_hash,
 					struct bkey_i *new_key,
+					unsigned commit_flags,
 					bool skip_triggers)
 {
 	struct bch_fs *c = trans->c;
@@ -2126,11 +2175,11 @@ static int __bch2_btree_node_update_key(struct btree_trans *trans,
 				_THIS_IP_);
 
 		BUG_ON(iter2.path->level != b->c.level);
-		BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
+		BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
 
 		btree_path_set_level_up(trans, iter2.path);
 
-		bch2_btree_path_check_sort(trans, iter2.path, 0);
+		trans->paths_sorted = false;
 
 		ret   = bch2_btree_iter_traverse(&iter2) ?:
 			bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
@@ -2151,12 +2200,7 @@ static int __bch2_btree_node_update_key(struct btree_trans *trans,
 		trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
 	}
 
-	ret = bch2_trans_commit(trans, NULL, NULL,
-				BTREE_INSERT_NOFAIL|
-				BTREE_INSERT_NOCHECK_RW|
-				BTREE_INSERT_USE_RESERVE|
-				BTREE_INSERT_JOURNAL_RECLAIM|
-				JOURNAL_WATERMARK_reserved);
+	ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
 	if (ret)
 		goto err;
 
@@ -2190,7 +2234,7 @@ err:
 
 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
 			       struct btree *b, struct bkey_i *new_key,
-			       bool skip_triggers)
+			       unsigned commit_flags, bool skip_triggers)
 {
 	struct bch_fs *c = trans->c;
 	struct btree *new_hash = NULL;
@@ -2211,19 +2255,17 @@ int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *ite
 	if (btree_ptr_hash_val(new_key) != b->hash_val) {
 		ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 		if (ret) {
-			bch2_trans_unlock(trans);
-			closure_sync(&cl);
-			ret = bch2_trans_relock(trans);
+			ret = drop_locks_do(trans, (closure_sync(&cl), 0));
 			if (ret)
 				return ret;
 		}
 
-		new_hash = bch2_btree_node_mem_alloc(c, false);
+		new_hash = bch2_btree_node_mem_alloc(trans, false);
 	}
 
 	path->intent_ref++;
-	ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
-					   new_key, skip_triggers);
+	ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
+					   commit_flags, skip_triggers);
 	--path->intent_ref;
 
 	if (new_hash) {
@@ -2241,7 +2283,7 @@ int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *ite
 
 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
 					struct btree *b, struct bkey_i *new_key,
-					bool skip_triggers)
+					unsigned commit_flags, bool skip_triggers)
 {
 	struct btree_iter iter;
 	int ret;
@@ -2262,7 +2304,8 @@ int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
 
 	BUG_ON(!btree_node_hashed(b));
 
-	ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
+	ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
+					 commit_flags, skip_triggers);
 out:
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
@@ -2281,8 +2324,9 @@ void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
 	bch2_btree_set_root_inmem(c, b);
 }
 
-void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
+static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
 {
+	struct bch_fs *c = trans->c;
 	struct closure cl;
 	struct btree *b;
 	int ret;
@@ -2294,7 +2338,7 @@ void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
 		closure_sync(&cl);
 	} while (ret);
 
-	b = bch2_btree_node_mem_alloc(c, false);
+	b = bch2_btree_node_mem_alloc(trans, false);
 	bch2_btree_cache_cannibalize_unlock(c);
 
 	set_btree_node_fake(b);
@@ -2323,6 +2367,12 @@ void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
 
 	six_unlock_write(&b->c.lock);
 	six_unlock_intent(&b->c.lock);
+	return 0;
+}
+
+void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
+{
+	bch2_trans_run(c, __bch2_btree_root_alloc(&trans, id));
 }
 
 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
@@ -2335,7 +2385,7 @@ void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
 		       as,
 		       as->mode,
 		       as->nodes_written,
-		       atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
+		       closure_nr_remaining(&as->cl),
 		       as->journal.seq);
 	mutex_unlock(&c->btree_interior_update_lock);
 }
@@ -2361,20 +2411,15 @@ bool bch2_btree_interior_updates_flush(struct bch_fs *c)
 	return ret;
 }
 
-void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
+void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
 {
-	struct btree_root *r;
-	struct jset_entry *entry;
+	struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
 
 	mutex_lock(&c->btree_root_lock);
 
-	vstruct_for_each(jset, entry)
-		if (entry->type == BCH_JSET_ENTRY_btree_root) {
-			r = &c->btree_roots[entry->btree_id];
-			r->level = entry->level;
-			r->alive = true;
-			bkey_copy(&r->key, &entry->start[0]);
-		}
+	r->level = entry->level;
+	r->alive = true;
+	bkey_copy(&r->key, &entry->start[0]);
 
 	mutex_unlock(&c->btree_root_lock);
 }
@@ -2394,15 +2439,15 @@ bch2_btree_roots_to_journal_entries(struct bch_fs *c,
 
 	mutex_lock(&c->btree_root_lock);
 
-	for (i = 0; i < BTREE_ID_NR; i++)
-		if (c->btree_roots[i].alive && !test_bit(i, &have)) {
-			journal_entry_set(end,
-					  BCH_JSET_ENTRY_btree_root,
-					  i, c->btree_roots[i].level,
-					  &c->btree_roots[i].key,
-					  c->btree_roots[i].key.u64s);
+	for (i = 0; i < btree_id_nr_alive(c); i++) {
+		struct btree_root *r = bch2_btree_id_root(c, i);
+
+		if (r->alive && !test_bit(i, &have)) {
+			journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
+					  i, r->level, &r->key, r->key.k.u64s);
 			end = vstruct_next(end);
 		}
+	}
 
 	mutex_unlock(&c->btree_root_lock);
 
@@ -2416,7 +2461,7 @@ void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
 	mempool_exit(&c->btree_interior_update_pool);
 }
 
-int bch2_fs_btree_interior_update_init(struct bch_fs *c)
+void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
 {
 	mutex_init(&c->btree_reserve_cache_lock);
 	INIT_LIST_HEAD(&c->btree_interior_update_list);
@@ -2424,11 +2469,20 @@ int bch2_fs_btree_interior_update_init(struct bch_fs *c)
 	mutex_init(&c->btree_interior_update_lock);
 	INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
 
+	INIT_LIST_HEAD(&c->pending_node_rewrites);
+	mutex_init(&c->pending_node_rewrites_lock);
+}
+
+int bch2_fs_btree_interior_update_init(struct bch_fs *c)
+{
 	c->btree_interior_update_worker =
 		alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
 	if (!c->btree_interior_update_worker)
-		return -ENOMEM;
+		return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
 
-	return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
-					 sizeof(struct btree_update));
+	if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
+				      sizeof(struct btree_update)))
+		return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
+
+	return 0;
 }