Update bcachefs sources to 3e0c5b0722 fixup! bcachefs: Btree write buffer

32 files changed, 1051 insertions, 298 deletions

diff --git a/.bcachefs_revision b/.bcachefs_revision
index b5ff7667..f6330b43 100644
--- a/.bcachefs_revision
+++ b/.bcachefs_revision
@@ -1 +1 @@
-dab31ca168df1c7a492a0a2a626ba6cd983b2a7a
+3e0c5b0722d7fccf0ba83435c5da8892b00c0fe0
diff --git a/include/trace/events/bcachefs.h b/include/trace/events/bcachefs.h
index ca5d6c8a..10e51bb3 100644
--- a/include/trace/events/bcachefs.h
+++ b/include/trace/events/bcachefs.h
@@ -1145,6 +1145,51 @@ TRACE_EVENT(trans_restart_key_cache_key_realloced,
 		  __entry->new_u64s)
 );
 
+DEFINE_EVENT(transaction_event,	trans_restart_write_buffer_flush,
+	TP_PROTO(struct btree_trans *trans,
+		 unsigned long caller_ip),
+	TP_ARGS(trans, caller_ip)
+);
+
+TRACE_EVENT(write_buffer_flush,
+	TP_PROTO(struct btree_trans *trans, size_t nr, size_t skipped, size_t fast, size_t size),
+	TP_ARGS(trans, nr, skipped, fast, size),
+
+	TP_STRUCT__entry(
+		__field(size_t,		nr		)
+		__field(size_t,		skipped		)
+		__field(size_t,		fast		)
+		__field(size_t,		size		)
+	),
+
+	TP_fast_assign(
+		__entry->nr	= nr;
+		__entry->skipped = skipped;
+		__entry->fast	= fast;
+		__entry->size	= size;
+	),
+
+	TP_printk("%zu/%zu skipped %zu fast %zu",
+		  __entry->nr, __entry->size, __entry->skipped, __entry->fast)
+);
+
+TRACE_EVENT(write_buffer_flush_slowpath,
+	TP_PROTO(struct btree_trans *trans, size_t nr, size_t size),
+	TP_ARGS(trans, nr, size),
+
+	TP_STRUCT__entry(
+		__field(size_t,		nr		)
+		__field(size_t,		size		)
+	),
+
+	TP_fast_assign(
+		__entry->nr	= nr;
+		__entry->size	= size;
+	),
+
+	TP_printk("%zu/%zu", __entry->nr, __entry->size)
+);
+
 #endif /* _TRACE_BCACHE_H */
 
 /* This part must be outside protection */
diff --git a/libbcachefs/alloc_background.c b/libbcachefs/alloc_background.c
index 6fd948f1..2db44365 100644
--- a/libbcachefs/alloc_background.c
+++ b/libbcachefs/alloc_background.c
@@ -414,14 +414,13 @@ void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c
 	prt_newline(out);
 	prt_printf(out, "io_time[WRITE]    %llu",	a->io_time[WRITE]);
 	prt_newline(out);
+	prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
+	prt_newline(out);
 
-	if (k.k->type == KEY_TYPE_alloc_v4) {
+	if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) {
 		struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
 		const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
 
-		prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a_raw.v));
-		prt_newline(out);
-
 		prt_printf(out, "backpointers:     %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
 		printbuf_indent_add(out, 2);
 
@@ -674,7 +673,7 @@ int bch2_bucket_gens_init(struct bch_fs *c)
 			ret = commit_do(&trans, NULL, NULL,
 					BTREE_INSERT_NOFAIL|
 					BTREE_INSERT_LAZY_RW,
-				__bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i));
+				__bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 			if (ret)
 				break;
 			have_bucket_gens_key = false;
@@ -694,7 +693,7 @@ int bch2_bucket_gens_init(struct bch_fs *c)
 		ret = commit_do(&trans, NULL, NULL,
 				BTREE_INSERT_NOFAIL|
 				BTREE_INSERT_LAZY_RW,
-			__bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i));
+			__bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 
 	bch2_trans_exit(&trans);
 
@@ -1934,7 +1933,7 @@ static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
 			freespace->k.p		= k.k->p;
 			freespace->k.size	= k.k->size;
 
-			ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace) ?:
+			ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace, 0) ?:
 				bch2_trans_commit(&trans, NULL, NULL,
 						  BTREE_INSERT_LAZY_RW|
 						  BTREE_INSERT_NOFAIL);
diff --git a/libbcachefs/backpointers.c b/libbcachefs/backpointers.c
index 0f8ffdf4..c269fc73 100644
--- a/libbcachefs/backpointers.c
+++ b/libbcachefs/backpointers.c
@@ -5,42 +5,11 @@
 #include "backpointers.h"
 #include "btree_cache.h"
 #include "btree_update.h"
+#include "btree_write_buffer.h"
 #include "error.h"
 
 #include <linux/mm.h>
 
-/*
- * Convert from pos in backpointer btree to pos of corresponding bucket in alloc
- * btree:
- */
-static inline struct bpos bp_pos_to_bucket(const struct bch_fs *c,
-					   struct bpos bp_pos)
-{
-	struct bch_dev *ca = bch_dev_bkey_exists(c, bp_pos.inode);
-	u64 bucket_sector = bp_pos.offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT;
-
-	return POS(bp_pos.inode, sector_to_bucket(ca, bucket_sector));
-}
-
-/*
- * Convert from pos in alloc btree + bucket offset to pos in backpointer btree:
- */
-static inline struct bpos bucket_pos_to_bp(const struct bch_fs *c,
-					   struct bpos bucket,
-					   u64 bucket_offset)
-{
-	struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
-	struct bpos ret;
-
-	ret = POS(bucket.inode,
-		  (bucket_to_sector(ca, bucket.offset) <<
-		   MAX_EXTENT_COMPRESS_RATIO_SHIFT) + bucket_offset);
-
-	BUG_ON(!bkey_eq(bucket, bp_pos_to_bucket(c, ret)));
-
-	return ret;
-}
-
 static bool extent_matches_bp(struct bch_fs *c,
 			      enum btree_id btree_id, unsigned level,
 			      struct bkey_s_c k,
@@ -200,51 +169,53 @@ err:
 	return ret;
 }
 
-int bch2_bucket_backpointer_del(struct btree_trans *trans,
-				struct bkey_i_alloc_v4 *a,
-				struct bch_backpointer bp,
-				struct bkey_s_c orig_k)
+bool bch2_bucket_backpointer_del(struct btree_trans *trans,
+				 struct bkey_i_alloc_v4 *a,
+				 struct bch_backpointer bp)
 {
-	struct bch_fs *c = trans->c;
 	struct bch_backpointer *bps = alloc_v4_backpointers(&a->v);
 	unsigned i, nr = BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v);
-	struct btree_iter bp_iter;
-	struct bkey_s_c k;
-	int ret;
 
 	for (i = 0; i < nr; i++) {
 		int cmp = backpointer_cmp(bps[i], bp) ?:
 			memcmp(&bps[i], &bp, sizeof(bp));
-		if (!cmp)
-			goto found;
+		if (!cmp) {
+			array_remove_item(bps, nr, i);
+			SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v, nr);
+			set_alloc_v4_u64s(a);
+			return true;
+		}
 		if (cmp >= 0)
 			break;
 	}
 
-	goto btree;
-found:
-	array_remove_item(bps, nr, i);
-	SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v, nr);
-	set_alloc_v4_u64s(a);
-	return 0;
-btree:
-	bch2_trans_iter_init(trans, &bp_iter, BTREE_ID_backpointers,
-			     bucket_pos_to_bp(c, a->k.p, bp.bucket_offset),
-			     BTREE_ITER_INTENT|
-			     BTREE_ITER_SLOTS|
-			     BTREE_ITER_WITH_UPDATES);
-	k = bch2_btree_iter_peek_slot(&bp_iter);
-	ret = bkey_err(k);
-	if (ret)
-		goto err;
+	return false;
+}
 
-	if (k.k->type != KEY_TYPE_backpointer ||
-	    memcmp(bkey_s_c_to_backpointer(k).v, &bp, sizeof(bp))) {
-		struct printbuf buf = PRINTBUF;
+static noinline int backpointer_mod_err(struct btree_trans *trans,
+					struct bch_backpointer bp,
+					struct bkey_s_c bp_k,
+					struct bkey_s_c orig_k,
+					bool insert)
+{
+	struct bch_fs *c = trans->c;
+	struct printbuf buf = PRINTBUF;
 
-		if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags))
-			goto err;
+	if (insert) {
+		prt_printf(&buf, "existing backpointer found when inserting ");
+		bch2_backpointer_to_text(&buf, &bp);
+		prt_newline(&buf);
+		printbuf_indent_add(&buf, 2);
+
+		prt_printf(&buf, "found ");
+		bch2_bkey_val_to_text(&buf, c, bp_k);
+		prt_newline(&buf);
+
+		prt_printf(&buf, "for ");
+		bch2_bkey_val_to_text(&buf, c, orig_k);
 
+		bch_err(c, "%s", buf.buf);
+	} else if (test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
 		prt_printf(&buf, "backpointer not found when deleting");
 		prt_newline(&buf);
 		printbuf_indent_add(&buf, 2);
@@ -254,85 +225,51 @@ btree:
 		prt_newline(&buf);
 
 		prt_printf(&buf, "got ");
-		bch2_bkey_val_to_text(&buf, c, k);
-		prt_newline(&buf);
-
-		prt_str(&buf, "alloc ");
-		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
+		bch2_bkey_val_to_text(&buf, c, bp_k);
 		prt_newline(&buf);
 
 		prt_printf(&buf, "for ");
 		bch2_bkey_val_to_text(&buf, c, orig_k);
 
 		bch_err(c, "%s", buf.buf);
-		bch2_inconsistent_error(c);
-		ret = -EIO;
-		printbuf_exit(&buf);
-		goto err;
 	}
 
-	ret = bch2_btree_delete_at(trans, &bp_iter, 0);
-err:
-	bch2_trans_iter_exit(trans, &bp_iter);
-	return ret;
+	printbuf_exit(&buf);
+
+	if (test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
+		bch2_inconsistent_error(c);
+		return -EIO;
+	} else {
+		return 0;
+	}
 }
 
-int bch2_bucket_backpointer_add(struct btree_trans *trans,
+int bch2_bucket_backpointer_mod_nowritebuffer(struct btree_trans *trans,
 				struct bkey_i_alloc_v4 *a,
 				struct bch_backpointer bp,
-				struct bkey_s_c orig_k)
+				struct bkey_s_c orig_k,
+				bool insert)
 {
 	struct bch_fs *c = trans->c;
-	struct bch_backpointer *bps = alloc_v4_backpointers(&a->v);
-	unsigned i, nr = BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v);
 	struct bkey_i_backpointer *bp_k;
 	struct btree_iter bp_iter;
 	struct bkey_s_c k;
 	int ret;
 
-	/* Check for duplicates: */
-	for (i = 0; i < nr; i++) {
-		int cmp = backpointer_cmp(bps[i], bp);
-		if (cmp >= 0)
-			break;
-	}
-
-	if ((i &&
-	     (bps[i - 1].bucket_offset +
-	      bps[i - 1].bucket_len > bp.bucket_offset)) ||
-	    (i < nr &&
-	     (bp.bucket_offset + bp.bucket_len > bps[i].bucket_offset))) {
-		struct printbuf buf = PRINTBUF;
-
-		prt_printf(&buf, "overlapping backpointer found when inserting ");
-		bch2_backpointer_to_text(&buf, &bp);
-		prt_newline(&buf);
-		printbuf_indent_add(&buf, 2);
-
-		prt_printf(&buf, "into ");
-		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
-		prt_newline(&buf);
-
-		prt_printf(&buf, "for ");
-		bch2_bkey_val_to_text(&buf, c, orig_k);
+	bp_k = bch2_trans_kmalloc_nomemzero(trans, sizeof(struct bkey_i_backpointer));
+	ret = PTR_ERR_OR_ZERO(bp_k);
+	if (ret)
+		return ret;
 
-		bch_err(c, "%s", buf.buf);
-		printbuf_exit(&buf);
-		if (test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
-			bch2_inconsistent_error(c);
-			return -EIO;
-		}
-	}
+	bkey_backpointer_init(&bp_k->k_i);
+	bp_k->k.p = bucket_pos_to_bp(c, a->k.p, bp.bucket_offset);
+	bp_k->v = bp;
 
-	if (nr < BCH_ALLOC_V4_NR_BACKPOINTERS_MAX) {
-		array_insert_item(bps, nr, i, bp);
-		SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v, nr);
-		set_alloc_v4_u64s(a);
-		return 0;
+	if (!insert) {
+		bp_k->k.type = KEY_TYPE_deleted;
+		set_bkey_val_u64s(&bp_k->k, 0);
 	}
 
-	/* Overflow: use backpointer btree */
-
 	bch2_trans_iter_init(trans, &bp_iter, BTREE_ID_backpointers,
 			     bucket_pos_to_bp(c, a->k.p, bp.bucket_offset),
 			     BTREE_ITER_INTENT|
@@ -343,37 +280,15 @@ int bch2_bucket_backpointer_add(struct btree_trans *trans,
 	if (ret)
 		goto err;
 
-	if (k.k->type) {
-		struct printbuf buf = PRINTBUF;
-
-		prt_printf(&buf, "existing btree backpointer key found when inserting ");
-		bch2_backpointer_to_text(&buf, &bp);
-		prt_newline(&buf);
-		printbuf_indent_add(&buf, 2);
-
-		prt_printf(&buf, "found ");
-		bch2_bkey_val_to_text(&buf, c, k);
-		prt_newline(&buf);
-
-		prt_printf(&buf, "for ");
-		bch2_bkey_val_to_text(&buf, c, orig_k);
-
-		bch_err(c, "%s", buf.buf);
-		printbuf_exit(&buf);
-		if (test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
-			bch2_inconsistent_error(c);
-			ret = -EIO;
+	if (insert
+	    ? k.k->type
+	    : (k.k->type != KEY_TYPE_backpointer ||
+	       memcmp(bkey_s_c_to_backpointer(k).v, &bp, sizeof(bp)))) {
+		ret = backpointer_mod_err(trans, bp, k, orig_k, insert);
+		if (ret)
 			goto err;
-		}
 	}
 
-	bp_k = bch2_bkey_alloc(trans, &bp_iter, backpointer);
-	ret = PTR_ERR_OR_ZERO(bp_k);
-	if (ret)
-		goto err;
-
-	bp_k->v = bp;
-
 	ret = bch2_trans_update(trans, &bp_iter, &bp_k->k_i, 0);
 err:
 	bch2_trans_iter_exit(trans, &bp_iter);
@@ -457,6 +372,9 @@ static void backpointer_not_found(struct btree_trans *trans,
 	struct bch_fs *c = trans->c;
 	struct printbuf buf = PRINTBUF;
 
+	if (likely(!bch2_backpointers_no_use_write_buffer))
+		return;
+
 	prt_printf(&buf, "backpointer doesn't match %s it points to:\n  ",
 		   thing_it_points_to);
 	prt_printf(&buf, "bucket: ");
@@ -512,27 +430,30 @@ struct bkey_s_c bch2_backpointer_get_key(struct btree_trans *trans,
 
 	bch2_trans_iter_exit(trans, iter);
 
-	if (bp.level) {
-		struct btree *b;
-
-		/*
-		 * If a backpointer for a btree node wasn't found, it may be
-		 * because it was overwritten by a new btree node that hasn't
-		 * been written out yet - backpointer_get_node() checks for
-		 * this:
-		 */
-		b = bch2_backpointer_get_node(trans, iter, bucket, bp_offset, bp);
-		if (!IS_ERR_OR_NULL(b))
-			return bkey_i_to_s_c(&b->key);
-
-		bch2_trans_iter_exit(trans, iter);
+	if (unlikely(bch2_backpointers_no_use_write_buffer)) {
+		if (bp.level) {
+			struct btree *b;
+
+			/*
+			 * If a backpointer for a btree node wasn't found, it may be
+			 * because it was overwritten by a new btree node that hasn't
+			 * been written out yet - backpointer_get_node() checks for
+			 * this:
+			 */
+			b = bch2_backpointer_get_node(trans, iter, bucket, bp_offset, bp);
+			if (!IS_ERR_OR_NULL(b))
+				return bkey_i_to_s_c(&b->key);
+
+			bch2_trans_iter_exit(trans, iter);
+
+			if (IS_ERR(b))
+				return bkey_s_c_err(PTR_ERR(b));
+			return bkey_s_c_null;
+		}
 
-		if (IS_ERR(b))
-			return bkey_s_c_err(PTR_ERR(b));
-		return bkey_s_c_null;
+		backpointer_not_found(trans, bucket, bp_offset, bp, k, "extent");
 	}
 
-	backpointer_not_found(trans, bucket, bp_offset, bp, k, "extent");
 	return bkey_s_c_null;
 }
 
@@ -633,7 +554,8 @@ static int check_bp_exists(struct btree_trans *trans,
 			   struct bch_backpointer bp,
 			   struct bkey_s_c orig_k,
 			   struct bpos bucket_start,
-			   struct bpos bucket_end)
+			   struct bpos bucket_end,
+			   struct bpos *last_flushed_pos)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter alloc_iter, bp_iter = { NULL };
@@ -677,8 +599,15 @@ static int check_bp_exists(struct btree_trans *trans,
 		goto err;
 
 	if (bp_k.k->type != KEY_TYPE_backpointer ||
-	    memcmp(bkey_s_c_to_backpointer(bp_k).v, &bp, sizeof(bp)))
+	    memcmp(bkey_s_c_to_backpointer(bp_k).v, &bp, sizeof(bp))) {
+		if (!bpos_eq(*last_flushed_pos, orig_k.k->p)) {
+			*last_flushed_pos = orig_k.k->p;
+			ret = bch2_btree_write_buffer_flush_sync(trans) ?:
+				-BCH_ERR_transaction_restart_write_buffer_flush;
+			goto out;
+		}
 		goto missing;
+	}
 out:
 err:
 fsck_err:
@@ -687,11 +616,12 @@ fsck_err:
 	printbuf_exit(&buf);
 	return ret;
 missing:
+
 	prt_printf(&buf, "missing backpointer for btree=%s l=%u ",
 	       bch2_btree_ids[bp.btree_id], bp.level);
 	bch2_bkey_val_to_text(&buf, c, orig_k);
-	prt_printf(&buf, "\nin alloc key ");
-	bch2_bkey_val_to_text(&buf, c, alloc_k);
+	prt_printf(&buf, "\nbp pos ");
+	bch2_bpos_to_text(&buf, bp_iter.pos);
 
 	if (c->sb.version < bcachefs_metadata_version_backpointers ||
 	    c->opts.reconstruct_alloc ||
@@ -699,8 +629,7 @@ missing:
 		struct bkey_i_alloc_v4 *a = bch2_alloc_to_v4_mut(trans, alloc_k);
 
 		ret   = PTR_ERR_OR_ZERO(a) ?:
-			bch2_bucket_backpointer_add(trans, a, bp, orig_k) ?:
-			bch2_trans_update(trans, &alloc_iter, &a->k_i, 0);
+			bch2_bucket_backpointer_mod(trans, a, bp, orig_k, true);
 	}
 
 	goto out;
@@ -709,7 +638,8 @@ missing:
 static int check_extent_to_backpointers(struct btree_trans *trans,
 					struct btree_iter *iter,
 					struct bpos bucket_start,
-					struct bpos bucket_end)
+					struct bpos bucket_end,
+					struct bpos *last_flushed_pos)
 {
 	struct bch_fs *c = trans->c;
 	struct bkey_ptrs_c ptrs;
@@ -736,7 +666,9 @@ static int check_extent_to_backpointers(struct btree_trans *trans,
 		bch2_extent_ptr_to_bp(c, iter->btree_id, iter->path->level,
 				      k, p, &bucket_pos, &bp);
 
-		ret = check_bp_exists(trans, bucket_pos, bp, k, bucket_start, bucket_end);
+		ret = check_bp_exists(trans, bucket_pos, bp, k,
+				      bucket_start, bucket_end,
+				      last_flushed_pos);
 		if (ret)
 			return ret;
 	}
@@ -747,7 +679,8 @@ static int check_extent_to_backpointers(struct btree_trans *trans,
 static int check_btree_root_to_backpointers(struct btree_trans *trans,
 					    enum btree_id btree_id,
 					    struct bpos bucket_start,
-					    struct bpos bucket_end)
+					    struct bpos bucket_end,
+					    struct bpos *last_flushed_pos)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
@@ -779,7 +712,9 @@ static int check_btree_root_to_backpointers(struct btree_trans *trans,
 		bch2_extent_ptr_to_bp(c, iter.btree_id, iter.path->level + 1,
 				      k, p, &bucket_pos, &bp);
 
-		ret = check_bp_exists(trans, bucket_pos, bp, k, bucket_start, bucket_end);
+		ret = check_bp_exists(trans, bucket_pos, bp, k,
+				      bucket_start, bucket_end,
+				      last_flushed_pos);
 		if (ret)
 			goto err;
 	}
@@ -859,6 +794,7 @@ static int bch2_check_extents_to_backpointers_pass(struct btree_trans *trans,
 {
 	struct btree_iter iter;
 	enum btree_id btree_id;
+	struct bpos last_flushed_pos = SPOS_MAX;
 	int ret = 0;
 
 	for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
@@ -874,7 +810,8 @@ static int bch2_check_extents_to_backpointers_pass(struct btree_trans *trans,
 					BTREE_INSERT_LAZY_RW|
 					BTREE_INSERT_NOFAIL,
 					check_extent_to_backpointers(trans, &iter,
-								bucket_start, bucket_end));
+								bucket_start, bucket_end,
+								&last_flushed_pos));
 			if (ret)
 				break;
 		} while (!bch2_btree_iter_advance(&iter));
@@ -888,7 +825,8 @@ static int bch2_check_extents_to_backpointers_pass(struct btree_trans *trans,
 				BTREE_INSERT_LAZY_RW|
 				BTREE_INSERT_NOFAIL,
 				check_btree_root_to_backpointers(trans, btree_id,
-							bucket_start, bucket_end));
+							bucket_start, bucket_end,
+							&last_flushed_pos));
 		if (ret)
 			break;
 	}
diff --git a/libbcachefs/backpointers.h b/libbcachefs/backpointers.h
index ac7b0932..ded1ab7f 100644
--- a/libbcachefs/backpointers.h
+++ b/libbcachefs/backpointers.h
@@ -2,6 +2,8 @@
 #ifndef _BCACHEFS_BACKPOINTERS_BACKGROUND_H
 #define _BCACHEFS_BACKPOINTERS_BACKGROUND_H
 
+#include "btree_iter.h"
+#include "btree_update.h"
 #include "buckets.h"
 #include "super.h"
 
@@ -19,6 +21,81 @@ void bch2_backpointer_swab(struct bkey_s);
 
 #define MAX_EXTENT_COMPRESS_RATIO_SHIFT		10
 
+/*
+ * Convert from pos in backpointer btree to pos of corresponding bucket in alloc
+ * btree:
+ */
+static inline struct bpos bp_pos_to_bucket(const struct bch_fs *c,
+					   struct bpos bp_pos)
+{
+	struct bch_dev *ca = bch_dev_bkey_exists(c, bp_pos.inode);
+	u64 bucket_sector = bp_pos.offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT;
+
+	return POS(bp_pos.inode, sector_to_bucket(ca, bucket_sector));
+}
+
+/*
+ * Convert from pos in alloc btree + bucket offset to pos in backpointer btree:
+ */
+static inline struct bpos bucket_pos_to_bp(const struct bch_fs *c,
+					   struct bpos bucket,
+					   u64 bucket_offset)
+{
+	struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
+	struct bpos ret;
+
+	ret = POS(bucket.inode,
+		  (bucket_to_sector(ca, bucket.offset) <<
+		   MAX_EXTENT_COMPRESS_RATIO_SHIFT) + bucket_offset);
+
+	BUG_ON(!bkey_eq(bucket, bp_pos_to_bucket(c, ret)));
+
+	return ret;
+}
+
+bool bch2_bucket_backpointer_del(struct btree_trans *,
+				struct bkey_i_alloc_v4 *,
+				struct bch_backpointer);
+
+int bch2_bucket_backpointer_mod_nowritebuffer(struct btree_trans *,
+				struct bkey_i_alloc_v4 *,
+				struct bch_backpointer, struct bkey_s_c, bool);
+
+static inline int bch2_bucket_backpointer_mod(struct btree_trans *trans,
+				struct bkey_i_alloc_v4 *a,
+				struct bch_backpointer bp,
+				struct bkey_s_c orig_k,
+				bool insert)
+{
+	struct bch_fs *c = trans->c;
+	struct bkey_i_backpointer *bp_k;
+	int ret;
+
+	if (!insert &&
+	    unlikely(BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v)) &&
+	    bch2_bucket_backpointer_del(trans, a, bp))
+		return 0;
+
+	if (unlikely(bch2_backpointers_no_use_write_buffer))
+		return bch2_bucket_backpointer_mod_nowritebuffer(trans, a, bp, orig_k, insert);
+
+	bp_k = bch2_trans_kmalloc_nomemzero(trans, sizeof(struct bkey_i_backpointer));
+	ret = PTR_ERR_OR_ZERO(bp_k);
+	if (ret)
+		return ret;
+
+	bkey_backpointer_init(&bp_k->k_i);
+	bp_k->k.p = bucket_pos_to_bp(c, a->k.p, bp.bucket_offset);
+	bp_k->v = bp;
+
+	if (!insert) {
+		bp_k->k.type = KEY_TYPE_deleted;
+		set_bkey_val_u64s(&bp_k->k, 0);
+	}
+
+	return bch2_trans_update_buffered(trans, BTREE_ID_backpointers, &bp_k->k_i);
+}
+
 static inline void bch2_extent_ptr_to_bp(struct bch_fs *c,
 			   enum btree_id btree_id, unsigned level,
 			   struct bkey_s_c k, struct extent_ptr_decoded p,
@@ -40,10 +117,6 @@ static inline void bch2_extent_ptr_to_bp(struct bch_fs *c,
 	};
 }
 
-int bch2_bucket_backpointer_del(struct btree_trans *, struct bkey_i_alloc_v4 *,
-				struct bch_backpointer, struct bkey_s_c);
-int bch2_bucket_backpointer_add(struct btree_trans *, struct bkey_i_alloc_v4 *,
-				struct bch_backpointer, struct bkey_s_c);
 int bch2_get_next_backpointer(struct btree_trans *, struct bpos, int,
 			      u64 *, struct bch_backpointer *, unsigned);
 struct bkey_s_c bch2_backpointer_get_key(struct btree_trans *, struct btree_iter *,
diff --git a/libbcachefs/bcachefs.h b/libbcachefs/bcachefs.h
index f5131733..c8620127 100644
--- a/libbcachefs/bcachefs.h
+++ b/libbcachefs/bcachefs.h
@@ -316,7 +316,10 @@ do {									\
 		"done in memory")					\
 	BCH_DEBUG_PARAM(verify_all_btree_replicas,			\
 		"When reading btree nodes, read all replicas and "	\
-		"compare them")
+		"compare them")						\
+	BCH_DEBUG_PARAM(backpointers_no_use_write_buffer,		\
+		"Don't use the write buffer for backpointers, enabling "\
+		"extra runtime checks")
 
 /* Parameters that should only be compiled in debug mode: */
 #define BCH_DEBUG_PARAMS_DEBUG()					\
@@ -393,6 +396,7 @@ enum bch_time_stats {
 
 #include "alloc_types.h"
 #include "btree_types.h"
+#include "btree_write_buffer_types.h"
 #include "buckets_types.h"
 #include "buckets_waiting_for_journal_types.h"
 #include "clock_types.h"
@@ -581,6 +585,7 @@ struct btree_transaction_stats {
 	struct bch2_time_stats	lock_hold_times;
 	struct mutex		lock;
 	unsigned		nr_max_paths;
+	unsigned		wb_updates_size;
 	unsigned		max_mem;
 	char			*max_paths_text;
 };
@@ -775,6 +780,9 @@ struct bch_fs {
 	struct workqueue_struct	*btree_interior_update_worker;
 	struct work_struct	btree_interior_update_work;
 
+	struct list_head	pending_node_rewrites;
+	struct mutex		pending_node_rewrites_lock;
+
 	/* btree_io.c: */
 	spinlock_t		btree_write_error_lock;
 	struct btree_write_stats {
@@ -795,6 +803,8 @@ struct bch_fs {
 	struct btree_key_cache	btree_key_cache;
 	unsigned		btree_key_cache_btrees;
 
+	struct btree_write_buffer btree_write_buffer;
+
 	struct workqueue_struct	*btree_update_wq;
 	struct workqueue_struct	*btree_io_complete_wq;
 	/* copygc needs its own workqueue for index updates.. */
diff --git a/libbcachefs/bcachefs_format.h b/libbcachefs/bcachefs_format.h
index ffd91373..7574eccd 100644
--- a/libbcachefs/bcachefs_format.h
+++ b/libbcachefs/bcachefs_format.h
@@ -1472,7 +1472,8 @@ struct bch_sb_field_disk_groups {
 	x(trans_traverse_all,				71)	\
 	x(transaction_commit,				72)	\
 	x(write_super,					73)	\
-	x(trans_restart_would_deadlock_recursion_limit,	74)
+	x(trans_restart_would_deadlock_recursion_limit,	74)	\
+	x(trans_restart_write_buffer_flush,		75)
 
 enum bch_persistent_counters {
 #define x(t, n, ...) BCH_COUNTER_##t,
@@ -1707,6 +1708,7 @@ LE64_BITMASK(BCH_SB_JOURNAL_FLUSH_DISABLED,struct bch_sb, flags[3], 62, 63);
 LE64_BITMASK(BCH_SB_JOURNAL_RECLAIM_DELAY,struct bch_sb, flags[4], 0, 32);
 LE64_BITMASK(BCH_SB_JOURNAL_TRANSACTION_NAMES,struct bch_sb, flags[4], 32, 33);
 LE64_BITMASK(BCH_SB_NOCOW,		struct bch_sb, flags[4], 33, 34);
+LE64_BITMASK(BCH_SB_WRITE_BUFFER_SIZE,	struct bch_sb, flags[4], 34, 54);
 
 /*
  * Features:
diff --git a/libbcachefs/btree_io.c b/libbcachefs/btree_io.c
index 19565b94..f9a3f131 100644
--- a/libbcachefs/btree_io.c
+++ b/libbcachefs/btree_io.c
@@ -1246,9 +1246,7 @@ start:
 	bio_put(&rb->bio);
 	printbuf_exit(&buf);
 
-	if (!btree_node_read_error(b) &&
-	    (saw_error ||
-	     btree_node_need_rewrite(b))) {
+	if (saw_error && !btree_node_read_error(b)) {
 		struct printbuf buf = PRINTBUF;
 
 		bch2_bpos_to_text(&buf, b->key.k.p);
diff --git a/libbcachefs/btree_iter.c b/libbcachefs/btree_iter.c
index 077d72bf..dc2b2a08 100644
--- a/libbcachefs/btree_iter.c
+++ b/libbcachefs/btree_iter.c
@@ -1374,6 +1374,7 @@ noinline __cold
 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
 {
 	struct btree_insert_entry *i;
+	struct btree_write_buffered_key *wb;
 
 	prt_printf(buf, "transaction updates for %s journal seq %llu",
 	       trans->fn, trans->journal_res.seq);
@@ -1398,6 +1399,17 @@ void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
 		prt_newline(buf);
 	}
 
+	trans_for_each_wb_update(trans, wb) {
+		prt_printf(buf, "update: btree=%s wb=1 %pS",
+		       bch2_btree_ids[wb->btree],
+		       (void *) i->ip_allocated);
+		prt_newline(buf);
+
+		prt_printf(buf, "  new ");
+		bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(&wb->k));
+		prt_newline(buf);
+	}
+
 	printbuf_indent_sub(buf, 2);
 }
 
@@ -2931,8 +2943,11 @@ void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, unsigned fn_
 			trans->mem_bytes = expected_mem_bytes;
 		}
 	}
-	if (s)
+
+	if (s) {
 		trans->nr_max_paths = s->nr_max_paths;
+		trans->wb_updates_size = s->wb_updates_size;
+	}
 
 	trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 	trans->srcu_lock_time	= jiffies;
diff --git a/libbcachefs/btree_types.h b/libbcachefs/btree_types.h
index 23e7f0ca..ad73cd2e 100644
--- a/libbcachefs/btree_types.h
+++ b/libbcachefs/btree_types.h
@@ -9,6 +9,7 @@
 //#include "bkey_methods.h"
 #include "buckets_types.h"
 #include "darray.h"
+#include "errcode.h"
 #include "journal_types.h"
 #include "replicas_types.h"
 
@@ -429,6 +430,8 @@ struct btree_trans {
 	u8			fn_idx;
 	u8			nr_sorted;
 	u8			nr_updates;
+	u8			nr_wb_updates;
+	u8			wb_updates_size;
 	bool			used_mempool:1;
 	bool			in_traverse_all:1;
 	bool			paths_sorted:1;
@@ -459,6 +462,7 @@ struct btree_trans {
 	u8			sorted[BTREE_ITER_MAX + 8];
 	struct btree_path	*paths;
 	struct btree_insert_entry *updates;
+	struct btree_write_buffered_key *wb_updates;
 
 	/* update path: */
 	struct btree_trans_commit_hook *hooks;
diff --git a/libbcachefs/btree_update.h b/libbcachefs/btree_update.h
index 673c3a78..771e4b23 100644
--- a/libbcachefs/btree_update.h
+++ b/libbcachefs/btree_update.h
@@ -15,6 +15,9 @@ bool bch2_btree_bset_insert_key(struct btree_trans *, struct btree_path *,
 				struct bkey_i *);
 void bch2_btree_add_journal_pin(struct bch_fs *, struct btree *, u64);
 
+void bch2_btree_insert_key_leaf(struct btree_trans *, struct btree_path *,
+				struct bkey_i *, u64);
+
 enum btree_insert_flags {
 	/* First two bits for journal watermark: */
 	__BTREE_INSERT_NOFAIL = 2,
@@ -55,7 +58,8 @@ int bch2_btree_delete_extent_at(struct btree_trans *, struct btree_iter *,
 				unsigned, unsigned);
 int bch2_btree_delete_at(struct btree_trans *, struct btree_iter *, unsigned);
 
-int __bch2_btree_insert(struct btree_trans *, enum btree_id, struct bkey_i *);
+int __bch2_btree_insert(struct btree_trans *, enum btree_id, struct bkey_i *,
+			enum btree_update_flags);
 int bch2_btree_insert(struct bch_fs *, enum btree_id, struct bkey_i *,
 		     struct disk_reservation *, u64 *, int flags);
 
@@ -77,6 +81,8 @@ int bch2_trans_update_extent(struct btree_trans *, struct btree_iter *,
 
 int __must_check bch2_trans_update(struct btree_trans *, struct btree_iter *,
 				   struct bkey_i *, enum btree_update_flags);
+int __must_check bch2_trans_update_buffered(struct btree_trans *,
+					    enum btree_id, struct bkey_i *);
 
 void bch2_trans_commit_hook(struct btree_trans *,
 			    struct btree_trans_commit_hook *);
@@ -142,6 +148,11 @@ static inline int bch2_trans_commit(struct btree_trans *trans,
 	     (_i) < (_trans)->updates + (_trans)->nr_updates;		\
 	     (_i)++)
 
+#define trans_for_each_wb_update(_trans, _i)				\
+	for ((_i) = (_trans)->wb_updates;				\
+	     (_i) < (_trans)->wb_updates + (_trans)->nr_wb_updates;	\
+	     (_i)++)
+
 static inline void bch2_trans_reset_updates(struct btree_trans *trans)
 {
 	struct btree_insert_entry *i;
@@ -151,6 +162,8 @@ static inline void bch2_trans_reset_updates(struct btree_trans *trans)
 
 	trans->extra_journal_res	= 0;
 	trans->nr_updates		= 0;
+	trans->nr_wb_updates		= 0;
+	trans->wb_updates		= NULL;
 	trans->hooks			= NULL;
 	trans->extra_journal_entries.nr	= 0;
 
diff --git a/libbcachefs/btree_update_interior.c b/libbcachefs/btree_update_interior.c
index a40a0f5c..172fd3b6 100644
--- a/libbcachefs/btree_update_interior.c
+++ b/libbcachefs/btree_update_interior.c
@@ -1997,6 +1997,7 @@ 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;
@@ -2056,15 +2057,10 @@ void async_btree_node_rewrite_work(struct work_struct *work)
 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
 {
 	struct async_btree_rewrite *a;
-
-	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
-		bch_err(c, "%s: error getting c->writes ref", __func__);
-		return;
-	}
+	int ret;
 
 	a = kmalloc(sizeof(*a), GFP_NOFS);
 	if (!a) {
-		bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
 		bch_err(c, "%s: error allocating memory", __func__);
 		return;
 	}
@@ -2074,11 +2070,63 @@ 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,
@@ -2416,6 +2464,9 @@ 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);
+
 	c->btree_interior_update_worker =
 		alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
 	if (!c->btree_interior_update_worker)
diff --git a/libbcachefs/btree_update_interior.h b/libbcachefs/btree_update_interior.h
index 2e6d220c..30e9c137 100644
--- a/libbcachefs/btree_update_interior.h
+++ b/libbcachefs/btree_update_interior.h
@@ -318,6 +318,9 @@ void bch2_journal_entries_to_btree_roots(struct bch_fs *, struct jset *);
 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
 					struct jset_entry *, struct jset_entry *);
 
+void bch2_do_pending_node_rewrites(struct bch_fs *);
+void bch2_free_pending_node_rewrites(struct bch_fs *);
+
 void bch2_fs_btree_interior_update_exit(struct bch_fs *);
 int bch2_fs_btree_interior_update_init(struct bch_fs *);
 
diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_update_leaf.c
index 6cbc0e5e..0195b13d 100644
--- a/libbcachefs/btree_update_leaf.c
+++ b/libbcachefs/btree_update_leaf.c
@@ -8,6 +8,7 @@
 #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"
@@ -100,9 +101,6 @@ inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
 {
 	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);
@@ -252,25 +250,26 @@ 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);
@@ -288,6 +287,12 @@ 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: */
@@ -594,6 +599,7 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 {
 	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;
@@ -638,6 +644,10 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 			marking = true;
 	}
 
+	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:
@@ -674,17 +684,25 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 	    bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 		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;
 	}
 
 	if (unlikely(trans->extra_journal_entries.nr)) {
@@ -697,10 +715,10 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 	}
 
 	if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
-		trans_for_each_update(trans, i) {
-			struct journal *j = &c->journal;
-			struct jset_entry *entry;
+		struct journal *j = &c->journal;
+		struct jset_entry *entry;
 
+		trans_for_each_update(trans, i) {
 			if (i->key_cache_already_flushed)
 				continue;
 
@@ -725,6 +743,14 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 			bkey_copy(&entry->start[0], i->k);
 		}
 
+		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 (trans->journal_seq)
 			*trans->journal_seq = trans->journal_res.seq;
 	}
@@ -742,6 +768,12 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 		}
 	}
 
+	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;
 }
 
@@ -769,7 +801,8 @@ static inline int trans_lock_write(struct btree_trans *trans)
 		if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 			return trans_lock_write_fail(trans, i);
 
-		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;
@@ -778,9 +811,13 @@ static inline int trans_lock_write(struct btree_trans *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
@@ -821,10 +858,11 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
-	struct printbuf buf = PRINTBUF;
 	int ret, u64s_delta = 0;
 
 #ifdef CONFIG_BCACHEFS_DEBUG
+	struct printbuf buf = PRINTBUF;
+
 	trans_for_each_update(trans, i) {
 		int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 
@@ -833,8 +871,8 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags
 			return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
 		btree_insert_entry_checks(trans, i);
 	}
-#endif
 	printbuf_exit(&buf);
+#endif
 
 	trans_for_each_update(trans, i) {
 		if (i->cached)
@@ -962,6 +1000,30 @@ int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 		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;
+
+		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;
@@ -1023,10 +1085,12 @@ 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;
 
@@ -1049,6 +1113,20 @@ int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 			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));
@@ -1089,6 +1167,9 @@ int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 			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,
@@ -1606,6 +1687,59 @@ 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)
 {
@@ -1614,7 +1748,8 @@ void bch2_trans_commit_hook(struct btree_trans *trans,
 }
 
 int __bch2_btree_insert(struct btree_trans *trans,
-			enum btree_id id, struct bkey_i *k)
+			enum btree_id id,
+			struct bkey_i *k, enum btree_update_flags flags)
 {
 	struct btree_iter iter;
 	int ret;
@@ -1622,7 +1757,7 @@ int __bch2_btree_insert(struct btree_trans *trans,
 	bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
 			     BTREE_ITER_INTENT);
 	ret   = bch2_btree_iter_traverse(&iter) ?:
-		bch2_trans_update(trans, &iter, k, 0);
+		bch2_trans_update(trans, &iter, k, flags);
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
@@ -1640,7 +1775,7 @@ 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_extent_at(struct btree_trans *trans, struct btree_iter *iter,
diff --git a/libbcachefs/btree_write_buffer.c b/libbcachefs/btree_write_buffer.c
new file mode 100644
index 00000000..05b755a0
--- /dev/null
+++ b/libbcachefs/btree_write_buffer.c
@@ -0,0 +1,328 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "btree_locking.h"
+#include "btree_update.h"
+#include "btree_update_interior.h"
+#include "btree_write_buffer.h"
+#include "error.h"
+#include "journal.h"
+#include "journal_reclaim.h"
+
+#include <linux/sort.h>
+
+static int btree_write_buffered_key_cmp(const void *_l, const void *_r)
+{
+	const struct btree_write_buffered_key *l = _l;
+	const struct btree_write_buffered_key *r = _r;
+
+	return  cmp_int(l->btree, r->btree) ?:
+		bpos_cmp(l->k.k.p, r->k.k.p) ?:
+		cmp_int(l->journal_seq, r->journal_seq) ?:
+		cmp_int(l->journal_offset, r->journal_offset);
+}
+
+static int btree_write_buffered_journal_cmp(const void *_l, const void *_r)
+{
+	const struct btree_write_buffered_key *l = _l;
+	const struct btree_write_buffered_key *r = _r;
+
+	return  cmp_int(l->journal_seq, r->journal_seq);
+}
+
+static int bch2_btree_write_buffer_flush_one(struct btree_trans *trans,
+					     struct btree_iter *iter,
+					     struct btree_write_buffered_key *wb,
+					     unsigned commit_flags,
+					     bool *write_locked,
+					     size_t *fast)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_path *path;
+	int ret;
+
+	ret = bch2_btree_iter_traverse(iter);
+	if (ret)
+		return ret;
+
+	path = iter->path;
+
+	if (!*write_locked) {
+		ret = bch2_btree_node_lock_write(trans, path, &path->l[0].b->c);
+		if (ret)
+			return ret;
+
+		bch2_btree_node_prep_for_write(trans, path, path->l[0].b);
+		*write_locked = true;
+	}
+
+	if (!bch2_btree_node_insert_fits(c, path->l[0].b, wb->k.k.u64s)) {
+		bch2_btree_node_unlock_write(trans, path, path->l[0].b);
+		*write_locked = false;
+		goto trans_commit;
+	}
+
+	bch2_btree_insert_key_leaf(trans, path, &wb->k, wb->journal_seq);
+	(*fast)++;
+	return 0;
+trans_commit:
+	return  bch2_trans_update(trans, iter, &wb->k, 0) ?:
+		bch2_trans_commit(trans, NULL, NULL,
+				  commit_flags|
+				  BTREE_INSERT_NOFAIL|
+				  BTREE_INSERT_JOURNAL_RECLAIM);
+}
+
+static union btree_write_buffer_state btree_write_buffer_switch(struct btree_write_buffer *wb)
+{
+	union btree_write_buffer_state old, new;
+	u64 v = READ_ONCE(wb->state.v);
+
+	do {
+		old.v = new.v = v;
+
+		new.nr = 0;
+		new.idx++;
+	} while ((v = atomic64_cmpxchg_acquire(&wb->state.counter, old.v, new.v)) != old.v);
+
+	while (old.idx == 0 ? wb->state.ref0 : wb->state.ref1)
+		cpu_relax();
+
+	return old;
+}
+
+int __bch2_btree_write_buffer_flush(struct btree_trans *trans, unsigned commit_flags,
+				    bool locked)
+{
+	struct bch_fs *c = trans->c;
+	struct journal *j = &c->journal;
+	struct btree_write_buffer *wb = &c->btree_write_buffer;
+	struct journal_entry_pin pin;
+	struct btree_write_buffered_key *i, *dst, *keys;
+	struct btree_iter iter = { NULL };
+	size_t nr = 0, skipped = 0, fast = 0;
+	bool write_locked = false;
+	union btree_write_buffer_state s;
+	int ret = 0;
+
+	memset(&pin, 0, sizeof(pin));
+
+	if (!locked && !mutex_trylock(&wb->flush_lock))
+		return 0;
+
+	bch2_journal_pin_copy(j, &pin, &wb->journal_pin, NULL);
+	bch2_journal_pin_drop(j, &wb->journal_pin);
+
+	s = btree_write_buffer_switch(wb);
+	keys = wb->keys[s.idx];
+	nr = s.nr;
+
+	/*
+	 * We first sort so that we can detect and skip redundant updates, and
+	 * then we attempt to flush in sorted btree order, as this is most
+	 * efficient.
+	 *
+	 * However, since we're not flushing in the order they appear in the
+	 * journal we won't be able to drop our journal pin until everything is
+	 * flushed - which means this could deadlock the journal, if we weren't
+	 * passing BTREE_INSERT_JORUNAL_RECLAIM. This causes the update to fail
+	 * if it would block taking a journal reservation.
+	 *
+	 * If that happens, we sort them by the order they appeared in the
+	 * journal - after dropping redundant entries - and then restart
+	 * flushing, this time dropping journal pins as we go.
+	 */
+
+	sort(keys, nr, sizeof(keys[0]),
+	     btree_write_buffered_key_cmp, NULL);
+
+	for (i = keys; i < keys + nr; i++) {
+		if (i + 1 < keys + nr &&
+		    i[0].btree == i[1].btree &&
+		    bpos_eq(i[0].k.k.p, i[1].k.k.p)) {
+			skipped++;
+			continue;
+		}
+
+		if (write_locked &&
+		    (iter.path->btree_id != i->btree ||
+		     bpos_gt(i->k.k.p, iter.path->l[0].b->key.k.p))) {
+			bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
+			write_locked = false;
+		}
+
+		if (!iter.path || iter.path->btree_id != i->btree) {
+			bch2_trans_iter_exit(trans, &iter);
+			bch2_trans_iter_init(trans, &iter, i->btree, i->k.k.p, BTREE_ITER_INTENT);
+		}
+
+		bch2_btree_iter_set_pos(&iter, i->k.k.p);
+		iter.path->preserve = false;
+
+		do {
+			ret = bch2_btree_write_buffer_flush_one(trans, &iter, i,
+						commit_flags, &write_locked, &fast);
+			if (!write_locked)
+				bch2_trans_begin(trans);
+		} while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
+
+		if (ret)
+			break;
+	}
+
+	if (write_locked)
+		bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
+	bch2_trans_iter_exit(trans, &iter);
+
+	trace_write_buffer_flush(trans, nr, skipped, fast, wb->size);
+
+	if (ret == -BCH_ERR_journal_reclaim_would_deadlock)
+		goto slowpath;
+
+	bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret));
+out:
+	bch2_journal_pin_drop(j, &pin);
+	mutex_unlock(&wb->flush_lock);
+	return ret;
+slowpath:
+	trace_write_buffer_flush_slowpath(trans, i - keys, nr);
+
+	dst = keys;
+	for (; i < keys + nr; i++) {
+		if (i + 1 < keys + nr &&
+		    i[0].btree == i[1].btree &&
+		    bpos_eq(i[0].k.k.p, i[1].k.k.p))
+			continue;
+
+		*dst = *i;
+		dst++;
+	}
+	nr = dst - keys;
+
+	sort(keys, nr, sizeof(keys[0]),
+	     btree_write_buffered_journal_cmp,
+	     NULL);
+
+	for (i = keys; i < keys + nr; i++) {
+		if (i->journal_seq > pin.seq) {
+			struct journal_entry_pin pin2;
+
+			memset(&pin2, 0, sizeof(pin2));
+
+			bch2_journal_pin_add(j, i->journal_seq, &pin2, NULL);
+			bch2_journal_pin_drop(j, &pin);
+			bch2_journal_pin_copy(j, &pin, &pin2, NULL);
+			bch2_journal_pin_drop(j, &pin2);
+		}
+
+		ret = commit_do(trans, NULL, NULL,
+				commit_flags|
+				BTREE_INSERT_NOFAIL|
+				BTREE_INSERT_JOURNAL_RECLAIM|
+				JOURNAL_WATERMARK_reserved,
+				__bch2_btree_insert(trans, i->btree, &i->k, 0));
+		if (bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret)))
+			break;
+	}
+
+	goto out;
+}
+
+int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
+{
+	bch2_trans_unlock(trans);
+	mutex_lock(&trans->c->btree_write_buffer.flush_lock);
+	return __bch2_btree_write_buffer_flush(trans, 0, true);
+}
+
+int bch2_btree_write_buffer_flush(struct btree_trans *trans)
+{
+	return __bch2_btree_write_buffer_flush(trans, 0, false);
+}
+
+static int bch2_btree_write_buffer_journal_flush(struct journal *j,
+				struct journal_entry_pin *_pin, u64 seq)
+{
+	struct bch_fs *c = container_of(j, struct bch_fs, journal);
+	struct btree_write_buffer *wb = &c->btree_write_buffer;
+
+	mutex_lock(&wb->flush_lock);
+
+	return bch2_trans_run(c,
+			__bch2_btree_write_buffer_flush(&trans, BTREE_INSERT_NOCHECK_RW, true));
+}
+
+static inline u64 btree_write_buffer_ref(int idx)
+{
+	return ((union btree_write_buffer_state) {
+		.ref0 = idx == 0,
+		.ref1 = idx == 1,
+	}).v;
+}
+
+int bch2_btree_insert_keys_write_buffer(struct btree_trans *trans)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_write_buffer *wb = &c->btree_write_buffer;
+	struct btree_write_buffered_key *i;
+	union btree_write_buffer_state old, new;
+	int ret = 0;
+	u64 v;
+
+	trans_for_each_wb_update(trans, i) {
+		EBUG_ON(i->k.k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
+
+		i->journal_seq		= trans->journal_res.seq;
+		i->journal_offset	= trans->journal_res.offset;
+	}
+
+	preempt_disable();
+	v = READ_ONCE(wb->state.v);
+	do {
+		old.v = new.v = v;
+
+		new.v += btree_write_buffer_ref(new.idx);
+		new.nr += trans->nr_wb_updates;
+		if (new.nr > wb->size) {
+			ret = -BCH_ERR_btree_insert_need_flush_buffer;
+			goto out;
+		}
+	} while ((v = atomic64_cmpxchg_acquire(&wb->state.counter, old.v, new.v)) != old.v);
+
+	memcpy(wb->keys[new.idx] + old.nr,
+	       trans->wb_updates,
+	       sizeof(trans->wb_updates[0]) * trans->nr_wb_updates);
+
+	bch2_journal_pin_add(&c->journal, trans->journal_res.seq, &wb->journal_pin,
+			     bch2_btree_write_buffer_journal_flush);
+
+	atomic64_sub_return_release(btree_write_buffer_ref(new.idx), &wb->state.counter);
+out:
+	preempt_enable();
+	return ret;
+}
+
+void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
+{
+	struct btree_write_buffer *wb = &c->btree_write_buffer;
+
+	BUG_ON(wb->state.nr && !bch2_journal_error(&c->journal));
+
+	kvfree(wb->keys[1]);
+	kvfree(wb->keys[0]);
+}
+
+int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
+{
+	struct btree_write_buffer *wb = &c->btree_write_buffer;
+
+	mutex_init(&wb->flush_lock);
+	wb->size = c->opts.btree_write_buffer_size;
+
+	wb->keys[0] = kvmalloc_array(wb->size, sizeof(*wb->keys[0]), GFP_KERNEL);
+	wb->keys[1] = kvmalloc_array(wb->size, sizeof(*wb->keys[1]), GFP_KERNEL);
+	if (!wb->keys[0] || !wb->keys[1])
+		return -ENOMEM;
+
+	return 0;
+}
diff --git a/libbcachefs/btree_write_buffer.h b/libbcachefs/btree_write_buffer.h
new file mode 100644
index 00000000..322df1c8
--- /dev/null
+++ b/libbcachefs/btree_write_buffer.h
@@ -0,0 +1,14 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_BTREE_WRITE_BUFFER_H
+#define _BCACHEFS_BTREE_WRITE_BUFFER_H
+
+int __bch2_btree_write_buffer_flush(struct btree_trans *, unsigned, bool);
+int bch2_btree_write_buffer_flush_sync(struct btree_trans *);
+int bch2_btree_write_buffer_flush(struct btree_trans *);
+
+int bch2_btree_insert_keys_write_buffer(struct btree_trans *);
+
+void bch2_fs_btree_write_buffer_exit(struct bch_fs *);
+int bch2_fs_btree_write_buffer_init(struct bch_fs *);
+
+#endif /* _BCACHEFS_BTREE_WRITE_BUFFER_H */
diff --git a/libbcachefs/btree_write_buffer_types.h b/libbcachefs/btree_write_buffer_types.h
new file mode 100644
index 00000000..99993ba7
--- /dev/null
+++ b/libbcachefs/btree_write_buffer_types.h
@@ -0,0 +1,44 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_BTREE_WRITE_BUFFER_TYPES_H
+#define _BCACHEFS_BTREE_WRITE_BUFFER_TYPES_H
+
+#include "journal_types.h"
+
+#define BTREE_WRITE_BUFERED_VAL_U64s_MAX	4
+#define BTREE_WRITE_BUFERED_U64s_MAX	(BKEY_U64s + BTREE_WRITE_BUFERED_VAL_U64s_MAX)
+
+struct btree_write_buffered_key {
+	u64			journal_seq;
+	unsigned		journal_offset;
+	enum btree_id		btree;
+	__BKEY_PADDED(k, BTREE_WRITE_BUFERED_VAL_U64s_MAX);
+};
+
+union btree_write_buffer_state {
+	struct {
+		atomic64_t	counter;
+	};
+
+	struct {
+		u64		v;
+	};
+
+	struct {
+		u64			nr:23;
+		u64			idx:1;
+		u64			ref0:20;
+		u64			ref1:20;
+	};
+};
+
+struct btree_write_buffer {
+	struct mutex			flush_lock;
+	struct journal_entry_pin	journal_pin;
+
+	union btree_write_buffer_state	state;
+	size_t				size;
+
+	struct btree_write_buffered_key	*keys[2];
+};
+
+#endif /* _BCACHEFS_BTREE_WRITE_BUFFER_TYPES_H */
diff --git a/libbcachefs/buckets.c b/libbcachefs/buckets.c
index ac0328c4..ebfbfd97 100644
--- a/libbcachefs/buckets.c
+++ b/libbcachefs/buckets.c
@@ -663,13 +663,14 @@ err:
 	return ret;
 }
 
-static int check_bucket_ref(struct bch_fs *c,
+static int check_bucket_ref(struct btree_trans *trans,
 			    struct bkey_s_c k,
 			    const struct bch_extent_ptr *ptr,
 			    s64 sectors, enum bch_data_type ptr_data_type,
 			    u8 b_gen, u8 bucket_data_type,
 			    u32 dirty_sectors, u32 cached_sectors)
 {
+	struct bch_fs *c = trans->c;
 	struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 	size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 	u16 bucket_sectors = !ptr->cached
@@ -756,9 +757,12 @@ static int check_bucket_ref(struct bch_fs *c,
 		ret = -EIO;
 		goto err;
 	}
-err:
+out:
 	printbuf_exit(&buf);
 	return ret;
+err:
+	bch2_dump_trans_updates(trans);
+	goto out;
 }
 
 static int mark_stripe_bucket(struct btree_trans *trans,
@@ -800,7 +804,7 @@ static int mark_stripe_bucket(struct btree_trans *trans,
 	bucket_lock(g);
 	old = *g;
 
-	ret = check_bucket_ref(c, k, ptr, sectors, data_type,
+	ret = check_bucket_ref(trans, k, ptr, sectors, data_type,
 			       g->gen, g->data_type,
 			       g->dirty_sectors, g->cached_sectors);
 	if (ret)
@@ -832,7 +836,7 @@ static int __mark_pointer(struct btree_trans *trans,
 	u32 *dst_sectors = !ptr->cached
 		? dirty_sectors
 		: cached_sectors;
-	int ret = check_bucket_ref(trans->c, k, ptr, sectors, ptr_data_type,
+	int ret = check_bucket_ref(trans, k, ptr, sectors, ptr_data_type,
 				   bucket_gen, *bucket_data_type,
 				   *dirty_sectors, *cached_sectors);
 
@@ -1220,7 +1224,8 @@ not_found:
 		new->k.p		= bkey_start_pos(p.k);
 		new->k.p.offset += *idx - start;
 		bch2_key_resize(&new->k, next_idx - *idx);
-		ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new->k_i);
+		ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new->k_i,
+					  BTREE_TRIGGER_NORUN);
 	}
 
 	*idx = next_idx;
@@ -1269,6 +1274,47 @@ int bch2_mark_reflink_p(struct btree_trans *trans,
 	return ret;
 }
 
+void bch2_trans_fs_usage_revert(struct btree_trans *trans,
+				struct replicas_delta_list *deltas)
+{
+	struct bch_fs *c = trans->c;
+	struct bch_fs_usage *dst;
+	struct replicas_delta *d, *top = (void *) deltas->d + deltas->used;
+	s64 added = 0;
+	unsigned i;
+
+	percpu_down_read(&c->mark_lock);
+	preempt_disable();
+	dst = fs_usage_ptr(c, trans->journal_res.seq, false);
+
+	/* revert changes: */
+	for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
+		switch (d->r.data_type) {
+		case BCH_DATA_btree:
+		case BCH_DATA_user:
+		case BCH_DATA_parity:
+			added += d->delta;
+		}
+		BUG_ON(__update_replicas(c, dst, &d->r, -d->delta));
+	}
+
+	dst->nr_inodes -= deltas->nr_inodes;
+
+	for (i = 0; i < BCH_REPLICAS_MAX; i++) {
+		added				-= deltas->persistent_reserved[i];
+		dst->reserved			-= deltas->persistent_reserved[i];
+		dst->persistent_reserved[i]	-= deltas->persistent_reserved[i];
+	}
+
+	if (added > 0) {
+		trans->disk_res->sectors += added;
+		this_cpu_add(*c->online_reserved, added);
+	}
+
+	preempt_enable();
+	percpu_up_read(&c->mark_lock);
+}
+
 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
 			      struct replicas_delta_list *deltas)
 {
@@ -1349,7 +1395,7 @@ need_mark:
 
 /* trans_mark: */
 
-static int bch2_trans_mark_pointer(struct btree_trans *trans,
+static inline int bch2_trans_mark_pointer(struct btree_trans *trans,
 				   enum btree_id btree_id, unsigned level,
 				   struct bkey_s_c k, struct extent_ptr_decoded p,
 				   unsigned flags)
@@ -1378,9 +1424,7 @@ static int bch2_trans_mark_pointer(struct btree_trans *trans,
 		goto err;
 
 	if (!p.ptr.cached) {
-		ret = insert
-			? bch2_bucket_backpointer_add(trans, a, bp, k)
-			: bch2_bucket_backpointer_del(trans, a, bp, k);
+		ret = bch2_bucket_backpointer_mod(trans, a, bp, k, insert);
 		if (ret)
 			goto err;
 	}
@@ -1518,7 +1562,7 @@ static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
 	if (IS_ERR(a))
 		return PTR_ERR(a);
 
-	ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
+	ret = check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
 			       a->v.gen, a->v.data_type,
 			       a->v.dirty_sectors, a->v.cached_sectors);
 	if (ret)
diff --git a/libbcachefs/buckets.h b/libbcachefs/buckets.h
index b6a1db76..3398c9c3 100644
--- a/libbcachefs/buckets.h
+++ b/libbcachefs/buckets.h
@@ -248,6 +248,7 @@ int bch2_trans_mark_inode(struct btree_trans *, enum btree_id, unsigned, struct
 int bch2_trans_mark_reservation(struct btree_trans *, enum btree_id, unsigned, struct bkey_s_c, struct bkey_i *, unsigned);
 int bch2_trans_mark_reflink_p(struct btree_trans *, enum btree_id, unsigned, struct bkey_s_c, struct bkey_i *, unsigned);
 
+void bch2_trans_fs_usage_revert(struct btree_trans *, struct replicas_delta_list *);
 int bch2_trans_fs_usage_apply(struct btree_trans *, struct replicas_delta_list *);
 
 int bch2_trans_mark_metadata_bucket(struct btree_trans *, struct bch_dev *,
diff --git a/libbcachefs/data_update.c b/libbcachefs/data_update.c
index 9df958b4..199e894e 100644
--- a/libbcachefs/data_update.c
+++ b/libbcachefs/data_update.c
@@ -309,9 +309,11 @@ void bch2_data_update_exit(struct data_update *update)
 		bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
 	const struct bch_extent_ptr *ptr;
 
-	bkey_for_each_ptr(ptrs, ptr)
+	bkey_for_each_ptr(ptrs, ptr) {
 		bch2_bucket_nocow_unlock(&c->nocow_locks,
-				       PTR_BUCKET_POS(c, ptr), 0);
+					 PTR_BUCKET_POS(c, ptr), 0);
+		percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
+	}
 
 	bch2_bkey_buf_exit(&update->k, c);
 	bch2_disk_reservation_put(c, &update->op.res);
@@ -411,6 +413,7 @@ int bch2_data_update_init(struct btree_trans *trans,
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 	const union bch_extent_entry *entry;
 	struct extent_ptr_decoded p;
+	const struct bch_extent_ptr *ptr;
 	unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 	unsigned int ptrs_locked = 0;
 	int ret;
@@ -436,6 +439,9 @@ int bch2_data_update_init(struct btree_trans *trans,
 	if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 		m->op.alloc_reserve = RESERVE_movinggc;
 
+	bkey_for_each_ptr(ptrs, ptr)
+		percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
+
 	i = 0;
 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 		bool locked;
@@ -508,7 +514,8 @@ err:
 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 		if ((1U << i) & ptrs_locked)
 			bch2_bucket_nocow_unlock(&c->nocow_locks,
-						PTR_BUCKET_POS(c, &p.ptr), 0);
+						 PTR_BUCKET_POS(c, &p.ptr), 0);
+		percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
 		i++;
 	}
 
diff --git a/libbcachefs/ec.c b/libbcachefs/ec.c
index 9f5a27da..3ebe699f 100644
--- a/libbcachefs/ec.c
+++ b/libbcachefs/ec.c
@@ -9,6 +9,7 @@
 #include "bset.h"
 #include "btree_gc.h"
 #include "btree_update.h"
+#include "btree_write_buffer.h"
 #include "buckets.h"
 #include "disk_groups.h"
 #include "ec.h"
@@ -921,13 +922,16 @@ static int ec_stripe_update_extents(struct bch_fs *c, struct ec_stripe_buf *s)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
+	ret = bch2_btree_write_buffer_flush(&trans);
+	if (ret)
+		goto err;
+
 	for (i = 0; i < nr_data; i++) {
 		ret = ec_stripe_update_bucket(&trans, s, i);
 		if (ret)
 			break;
 	}
-
-
+err:
 	bch2_trans_exit(&trans);
 
 	return ret;
diff --git a/libbcachefs/errcode.h b/libbcachefs/errcode.h
index b8c24f51..6129af61 100644
--- a/libbcachefs/errcode.h
+++ b/libbcachefs/errcode.h
@@ -42,6 +42,7 @@
 	x(BCH_ERR_transaction_restart,	transaction_restart_key_cache_realloced)\
 	x(BCH_ERR_transaction_restart,	transaction_restart_journal_preres_get)	\
 	x(BCH_ERR_transaction_restart,	transaction_restart_split_race)		\
+	x(BCH_ERR_transaction_restart,	transaction_restart_write_buffer_flush)	\
 	x(BCH_ERR_transaction_restart,	transaction_restart_nested)		\
 	x(0,				no_btree_node)				\
 	x(BCH_ERR_no_btree_node,	no_btree_node_relock)			\
@@ -58,6 +59,7 @@
 	x(BCH_ERR_btree_insert_fail,	btree_insert_need_mark_replicas)	\
 	x(BCH_ERR_btree_insert_fail,	btree_insert_need_journal_res)		\
 	x(BCH_ERR_btree_insert_fail,	btree_insert_need_journal_reclaim)	\
+	x(BCH_ERR_btree_insert_fail,	btree_insert_need_flush_buffer)		\
 	x(0,				backpointer_to_overwritten_btree_node)	\
 	x(0,				lock_fail_root_changed)			\
 	x(0,				journal_reclaim_would_deadlock)		\
diff --git a/libbcachefs/fsck.c b/libbcachefs/fsck.c
index 5887d781..52bb00b5 100644
--- a/libbcachefs/fsck.c
+++ b/libbcachefs/fsck.c
@@ -1797,7 +1797,8 @@ static int check_root_trans(struct btree_trans *trans)
 		ret = commit_do(trans, NULL, NULL,
 				      BTREE_INSERT_NOFAIL|
 				      BTREE_INSERT_LAZY_RW,
-			__bch2_btree_insert(trans, BTREE_ID_subvolumes, &root_subvol.k_i));
+			__bch2_btree_insert(trans, BTREE_ID_subvolumes,
+					    &root_subvol.k_i, 0));
 		if (ret) {
 			bch_err(c, "error writing root subvol: %s", bch2_err_str(ret));
 			goto err;
diff --git a/libbcachefs/io.c b/libbcachefs/io.c
index c0c33f78..64925db2 100644
--- a/libbcachefs/io.c
+++ b/libbcachefs/io.c
@@ -684,7 +684,7 @@ void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
 
 			bio_set_dev(&n->bio, ca->disk_sb.bdev);
 
-			if (IS_ENABLED(CONFIG_BCACHEFS_NO_IO) && type != BCH_DATA_btree) {
+			if (type != BCH_DATA_btree && unlikely(c->opts.no_data_io)) {
 				bio_endio(&n->bio);
 				continue;
 			}
@@ -2356,8 +2356,7 @@ static void __bch2_read_endio(struct work_struct *work)
 	}
 
 	csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
-	if (bch2_crc_cmp(csum, rbio->pick.crc.csum) &&
-	    !IS_ENABLED(CONFIG_BCACHEFS_NO_IO))
+	if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
 		goto csum_err;
 
 	/*
@@ -2808,7 +2807,7 @@ get_bio:
 			     bio_sectors(&rbio->bio));
 		bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
 
-		if (IS_ENABLED(CONFIG_BCACHEFS_NO_IO)) {
+		if (unlikely(c->opts.no_data_io)) {
 			if (likely(!(flags & BCH_READ_IN_RETRY)))
 				bio_endio(&rbio->bio);
 		} else {
diff --git a/libbcachefs/journal.c b/libbcachefs/journal.c
index 6214a919..957eeece 100644
--- a/libbcachefs/journal.c
+++ b/libbcachefs/journal.c
@@ -769,6 +769,7 @@ static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 	if (c) {
 		bch2_journal_flush_all_pins(&c->journal);
 		bch2_journal_block(&c->journal);
+		mutex_lock(&c->sb_lock);
 	}
 
 	bu		= kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
@@ -849,6 +850,9 @@ static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 	if (!new_fs)
 		spin_unlock(&c->journal.lock);
 
+	if (ja->nr != old_nr && !new_fs)
+		bch2_write_super(c);
+
 	if (c)
 		bch2_journal_unblock(&c->journal);
 
@@ -868,6 +872,9 @@ static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 		}
 	}
 err:
+	if (c)
+		mutex_unlock(&c->sb_lock);
+
 	if (ob && !new_fs)
 		for (i = 0; i < nr_got; i++)
 			bch2_open_bucket_put(c, ob[i]);
@@ -893,7 +900,6 @@ int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
 {
 	struct journal_device *ja = &ca->journal;
 	struct closure cl;
-	unsigned current_nr;
 	int ret = 0;
 
 	/* don't handle reducing nr of buckets yet: */
@@ -902,44 +908,44 @@ int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
 
 	closure_init_stack(&cl);
 
-	while (ja->nr != nr && (ret == 0 || ret == -BCH_ERR_bucket_alloc_blocked)) {
+	while (ja->nr != nr) {
 		struct disk_reservation disk_res = { 0, 0 };
 
-		closure_sync(&cl);
-
-		mutex_lock(&c->sb_lock);
-		current_nr = ja->nr;
-
 		/*
 		 * note: journal buckets aren't really counted as _sectors_ used yet, so
 		 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
 		 * when space used goes up without a reservation - but we do need the
 		 * reservation to ensure we'll actually be able to allocate:
+		 *
+		 * XXX: that's not right, disk reservations only ensure a
+		 * filesystem-wide allocation will succeed, this is a device
+		 * specific allocation - we can hang here:
 		 */
 
 		ret = bch2_disk_reservation_get(c, &disk_res,
 						bucket_to_sector(ca, nr - ja->nr), 1, 0);
-		if (ret) {
-			mutex_unlock(&c->sb_lock);
-			return ret;
-		}
+		if (ret)
+			break;
 
 		ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
 
 		bch2_disk_reservation_put(c, &disk_res);
 
-		if (ja->nr != current_nr)
-			bch2_write_super(c);
-		mutex_unlock(&c->sb_lock);
+		closure_sync(&cl);
+
+		if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
+			break;
 	}
 
+	if (ret)
+		bch_err(c, "%s: err %s", __func__, bch2_err_str(ret));
+
 	return ret;
 }
 
 int bch2_dev_journal_alloc(struct bch_dev *ca)
 {
 	unsigned nr;
-	int ret;
 
 	if (dynamic_fault("bcachefs:add:journal_alloc"))
 		return -ENOMEM;
@@ -956,15 +962,7 @@ int bch2_dev_journal_alloc(struct bch_dev *ca)
 		     min(1 << 13,
 			 (1 << 24) / ca->mi.bucket_size));
 
-	if (ca->fs)
-		mutex_lock(&ca->fs->sb_lock);
-
-	ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
-
-	if (ca->fs)
-		mutex_unlock(&ca->fs->sb_lock);
-
-	return ret;
+	return __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
 }
 
 /* startup/shutdown: */
diff --git a/libbcachefs/move.c b/libbcachefs/move.c
index 8321563d..7e7e9042 100644
--- a/libbcachefs/move.c
+++ b/libbcachefs/move.c
@@ -8,6 +8,7 @@
 #include "btree_gc.h"
 #include "btree_update.h"
 #include "btree_update_interior.h"
+#include "btree_write_buffer.h"
 #include "disk_groups.h"
 #include "ec.h"
 #include "errcode.h"
@@ -680,10 +681,19 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 			bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 	bch2_trans_iter_exit(&trans, &iter);
 
-	if (!ret) {
-		a = bch2_alloc_to_v4(k, &a_convert);
-		dirty_sectors = a->dirty_sectors;
-		bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
+	if (ret) {
+		bch_err(c, "%s: error looking up alloc key: %s", __func__, bch2_err_str(ret));
+		goto err;
+	}
+
+	a = bch2_alloc_to_v4(k, &a_convert);
+	dirty_sectors = a->dirty_sectors;
+	bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
+
+	ret = bch2_btree_write_buffer_flush(&trans);
+	if (ret) {
+		bch_err(c, "%s: error flushing btree write buffer: %s", __func__, bch2_err_str(ret));
+		goto err;
 	}
 
 	while (!(ret = move_ratelimit(&trans, ctxt))) {
@@ -712,7 +722,7 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 			if (ret)
 				goto err;
 			if (!k.k)
-				continue;
+				goto next;
 
 			bch2_bkey_buf_reassemble(&sk, c, k);
 			k = bkey_i_to_s_c(sk.k);
@@ -763,7 +773,7 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 			if (ret)
 				goto err;
 			if (!b)
-				continue;
+				goto next;
 
 			ret = bch2_btree_node_rewrite(&trans, &iter, b, 0);
 			bch2_trans_iter_exit(&trans, &iter);
@@ -779,7 +789,7 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 			atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 			atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 		}
-
+next:
 		bp_offset++;
 	}
 
diff --git a/libbcachefs/opts.h b/libbcachefs/opts.h
index 9d4bbec7..304718a0 100644
--- a/libbcachefs/opts.h
+++ b/libbcachefs/opts.h
@@ -206,6 +206,11 @@ enum opt_type {
 	  OPT_BOOL(),							\
 	  BCH2_NO_SB_OPT,		true,				\
 	  NULL,		"Stash pointer to in memory btree node in btree ptr")\
+	x(btree_write_buffer_size, u32,					\
+	  OPT_FS|OPT_MOUNT,						\
+	  OPT_UINT(16, (1U << 20) - 1),					\
+	  BCH2_NO_SB_OPT,		1U << 13,			\
+	  NULL,		"Number of btree write buffer entries")		\
 	x(gc_reserve_percent,		u8,				\
 	  OPT_FS|OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME,			\
 	  OPT_UINT(5, 21),						\
@@ -399,6 +404,12 @@ enum opt_type {
 	  NULL,		"Nocow mode: Writes will be done in place when possible.\n"\
 			"Snapshots and reflink will still caused writes to be COW\n"\
 			"Implicitly disables data checksumming, compression and encryption")\
+	x(no_data_io,			u8,				\
+	  OPT_FS|OPT_MOUNT,						\
+	  OPT_BOOL(),							\
+	  BCH2_NO_SB_OPT,			false,			\
+	  NULL,		"Skip submit_bio() for data reads and writes, "	\
+			"for performance testing purposes")		\
 	x(fs_size,			u64,				\
 	  OPT_DEVICE,							\
 	  OPT_UINT(0, S64_MAX),						\
diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index 8a78377b..178f0642 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -974,9 +974,15 @@ static int read_btree_roots(struct bch_fs *c)
 		}
 	}
 
-	for (i = 0; i < BTREE_ID_NR; i++)
-		if (!c->btree_roots[i].b)
+	for (i = 0; i < BTREE_ID_NR; i++) {
+		struct btree_root *r = &c->btree_roots[i];
+
+		if (!r->b) {
+			r->alive = false;
+			r->level = 0;
 			bch2_btree_root_alloc(c, i);
+		}
+	}
 fsck_err:
 	return ret;
 }
diff --git a/libbcachefs/replicas.c b/libbcachefs/replicas.c
index 482bedf4..3bff2195 100644
--- a/libbcachefs/replicas.c
+++ b/libbcachefs/replicas.c
@@ -304,13 +304,6 @@ static int replicas_table_update(struct bch_fs *c,
 					sizeof(u64), GFP_KERNEL)))
 			goto err;
 
-	memset(new_usage, 0, sizeof(new_usage));
-
-	for (i = 0; i < ARRAY_SIZE(new_usage); i++)
-		if (!(new_usage[i] = __alloc_percpu_gfp(bytes,
-					sizeof(u64), GFP_KERNEL)))
-			goto err;
-
 	if (!(new_base = kzalloc(bytes, GFP_KERNEL)) ||
 	    !(new_scratch  = kmalloc(scratch_bytes, GFP_KERNEL)) ||
 	    (c->usage_gc &&
diff --git a/libbcachefs/super.c b/libbcachefs/super.c
index 08bfed1b..8bed1185 100644
--- a/libbcachefs/super.c
+++ b/libbcachefs/super.c
@@ -16,6 +16,7 @@
 #include "btree_key_cache.h"
 #include "btree_update_interior.h"
 #include "btree_io.h"
+#include "btree_write_buffer.h"
 #include "buckets_waiting_for_journal.h"
 #include "chardev.h"
 #include "checksum.h"
@@ -418,6 +419,7 @@ static int __bch2_fs_read_write(struct bch_fs *c, bool early)
 	bch2_do_discards(c);
 	bch2_do_invalidates(c);
 	bch2_do_stripe_deletes(c);
+	bch2_do_pending_node_rewrites(c);
 	return 0;
 err:
 	__bch2_fs_read_only(c);
@@ -446,6 +448,7 @@ static void __bch2_fs_free(struct bch_fs *c)
 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
 		bch2_time_stats_exit(&c->times[i]);
 
+	bch2_free_pending_node_rewrites(c);
 	bch2_fs_counters_exit(c);
 	bch2_fs_snapshots_exit(c);
 	bch2_fs_quota_exit(c);
@@ -465,6 +468,7 @@ static void __bch2_fs_free(struct bch_fs *c)
 	bch2_fs_compress_exit(c);
 	bch2_journal_keys_free(&c->journal_keys);
 	bch2_journal_entries_free(c);
+	bch2_fs_btree_write_buffer_exit(c);
 	percpu_free_rwsem(&c->mark_lock);
 	free_percpu(c->online_reserved);
 
@@ -817,6 +821,7 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	    bch2_fs_btree_iter_init(c) ?:
 	    bch2_fs_btree_interior_update_init(c) ?:
 	    bch2_fs_buckets_waiting_for_journal_init(c) ?:
+	    bch2_fs_btree_write_buffer_init(c) ?:
 	    bch2_fs_subvolumes_init(c) ?:
 	    bch2_fs_io_init(c) ?:
 	    bch2_fs_nocow_locking_init(c) ?:
diff --git a/libbcachefs/sysfs.c b/libbcachefs/sysfs.c
index ebd10cd5..b981c87e 100644
--- a/libbcachefs/sysfs.c
+++ b/libbcachefs/sysfs.c
@@ -473,14 +473,14 @@ SHOW(bch2_fs)
 	if (attr == &sysfs_data_jobs)
 		data_progress_to_text(out, c);
 
-	if (attr == &sysfs_nocow_lock_table)
-		bch2_nocow_locks_to_text(out, &c->nocow_locks);
-
 #ifdef BCH_WRITE_REF_DEBUG
 	if (attr == &sysfs_write_refs)
 		bch2_write_refs_to_text(out, c);
 #endif
 
+	if (attr == &sysfs_nocow_lock_table)
+		bch2_nocow_locks_to_text(out, &c->nocow_locks);
+
 	return 0;
 }
 
@@ -664,10 +664,10 @@ struct attribute *bch2_fs_internal_files[] = {
 	&sysfs_stripes_heap,
 	&sysfs_open_buckets,
 	&sysfs_write_points,
-	&sysfs_nocow_lock_table,
 #ifdef BCH_WRITE_REF_DEBUG
 	&sysfs_write_refs,
 #endif
+	&sysfs_nocow_lock_table,
 	&sysfs_io_timers_read,
 	&sysfs_io_timers_write,
 
diff --git a/libbcachefs/tests.c b/libbcachefs/tests.c
index 80fce1c9..d352821d 100644
--- a/libbcachefs/tests.c
+++ b/libbcachefs/tests.c
@@ -592,7 +592,7 @@ static int rand_insert(struct bch_fs *c, u64 nr)
 		k.k.p.snapshot = U32_MAX;
 
 		ret = commit_do(&trans, NULL, NULL, 0,
-			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k.k_i));
+			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k.k_i, 0));
 		if (ret) {
 			bch_err(c, "%s(): error %s", __func__, bch2_err_str(ret));
 			break;
@@ -621,14 +621,14 @@ static int rand_insert_multi(struct bch_fs *c, u64 nr)
 		}
 
 		ret = commit_do(&trans, NULL, NULL, 0,
-			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[0].k_i) ?:
-			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[1].k_i) ?:
-			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[2].k_i) ?:
-			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[3].k_i) ?:
-			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[4].k_i) ?:
-			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[5].k_i) ?:
-			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[6].k_i) ?:
-			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[7].k_i));
+			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[0].k_i, 0) ?:
+			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[1].k_i, 0) ?:
+			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[2].k_i, 0) ?:
+			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[3].k_i, 0) ?:
+			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[4].k_i, 0) ?:
+			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[5].k_i, 0) ?:
+			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[6].k_i, 0) ?:
+			__bch2_btree_insert(&trans, BTREE_ID_xattrs, &k[7].k_i, 0));
 		if (ret) {
 			bch_err(c, "%s(): error %s", __func__, bch2_err_str(ret));
 			break;