X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Frecovery.c;h=ca92fe84c248a25f5863046bdc038f265274e20d;hb=70f2681838973f20a918b518f82d382b3556963d;hp=d24cef2bf1aa3252b611e1fee6e84a41dc2b4964;hpb=dd707c9bfdc13fd90856576ddbb383efccef751c;p=bcachefs-tools-debian

diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index d24cef2..ca92fe8 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
+#include "bkey_buf.h"
 #include "alloc_background.h"
 #include "btree_gc.h"
 #include "btree_update.h"
@@ -15,9 +16,12 @@
 #include "journal_io.h"
 #include "journal_reclaim.h"
 #include "journal_seq_blacklist.h"
+#include "lru.h"
+#include "move.h"
 #include "quota.h"
 #include "recovery.h"
 #include "replicas.h"
+#include "subvolume.h"
 #include "super-io.h"
 
 #include <linux/sort.h>
@@ -31,86 +35,246 @@ static void drop_alloc_keys(struct journal_keys *keys)
 	size_t src, dst;
 
 	for (src = 0, dst = 0; src < keys->nr; src++)
-		if (keys->d[src].btree_id != BTREE_ID_ALLOC)
+		if (keys->d[src].btree_id != BTREE_ID_alloc)
 			keys->d[dst++] = keys->d[src];
 
 	keys->nr = dst;
 }
 
+/*
+ * Btree node pointers have a field to stack a pointer to the in memory btree
+ * node; we need to zero out this field when reading in btree nodes, or when
+ * reading in keys from the journal:
+ */
+static void zero_out_btree_mem_ptr(struct journal_keys *keys)
+{
+	struct journal_key *i;
+
+	for (i = keys->d; i < keys->d + keys->nr; i++)
+		if (i->k->k.type == KEY_TYPE_btree_ptr_v2)
+			bkey_i_to_btree_ptr_v2(i->k)->v.mem_ptr = 0;
+}
+
 /* iterate over keys read from the journal: */
 
-static struct journal_key *journal_key_search(struct journal_keys *journal_keys,
-					      enum btree_id id, unsigned level,
-					      struct bpos pos)
+static int __journal_key_cmp(enum btree_id	l_btree_id,
+			     unsigned		l_level,
+			     struct bpos	l_pos,
+			     const struct journal_key *r)
+{
+	return (cmp_int(l_btree_id,	r->btree_id) ?:
+		cmp_int(l_level,	r->level) ?:
+		bpos_cmp(l_pos,	r->k->k.p));
+}
+
+static int journal_key_cmp(const struct journal_key *l, const struct journal_key *r)
+{
+	return __journal_key_cmp(l->btree_id, l->level, l->k->k.p, r);
+}
+
+size_t bch2_journal_key_search(struct journal_keys *journal_keys,
+			       enum btree_id id, unsigned level,
+			       struct bpos pos)
 {
 	size_t l = 0, r = journal_keys->nr, m;
 
 	while (l < r) {
 		m = l + ((r - l) >> 1);
-		if ((cmp_int(id,	journal_keys->d[m].btree_id) ?:
-		     cmp_int(level,	journal_keys->d[m].level) ?:
-		     bkey_cmp(pos,	journal_keys->d[m].k->k.p)) > 0)
+		if (__journal_key_cmp(id, level, pos, &journal_keys->d[m]) > 0)
 			l = m + 1;
 		else
 			r = m;
 	}
 
 	BUG_ON(l < journal_keys->nr &&
-	       (cmp_int(id,	journal_keys->d[l].btree_id) ?:
-		cmp_int(level,	journal_keys->d[l].level) ?:
-		bkey_cmp(pos,	journal_keys->d[l].k->k.p)) > 0);
+	       __journal_key_cmp(id, level, pos, &journal_keys->d[l]) > 0);
 
 	BUG_ON(l &&
-	       (cmp_int(id,	journal_keys->d[l - 1].btree_id) ?:
-		cmp_int(level,	journal_keys->d[l - 1].level) ?:
-		bkey_cmp(pos,	journal_keys->d[l - 1].k->k.p)) <= 0);
+	       __journal_key_cmp(id, level, pos, &journal_keys->d[l - 1]) <= 0);
+
+	return l;
+}
+
+struct bkey_i *bch2_journal_keys_peek(struct bch_fs *c, enum btree_id btree_id,
+				      unsigned level, struct bpos pos)
+{
+	struct journal_keys *keys = &c->journal_keys;
+	struct journal_key *end = keys->d + keys->nr;
+	struct journal_key *k = keys->d +
+		bch2_journal_key_search(keys, btree_id, level, pos);
+
+	while (k < end && k->overwritten)
+		k++;
+
+	if (k < end &&
+	    k->btree_id	== btree_id &&
+	    k->level	== level)
+		return k->k;
+	return NULL;
+}
+
+static void journal_iter_fix(struct bch_fs *c, struct journal_iter *iter, unsigned idx)
+{
+	struct bkey_i *n = iter->keys->d[idx].k;
+	struct btree_and_journal_iter *biter =
+		container_of(iter, struct btree_and_journal_iter, journal);
+
+	if (iter->idx > idx ||
+	    (iter->idx == idx &&
+	     biter->last &&
+	     bpos_cmp(n->k.p, biter->unpacked.p) <= 0))
+		iter->idx++;
+}
+
+int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
+				 unsigned level, struct bkey_i *k)
+{
+	struct journal_key n = {
+		.btree_id	= id,
+		.level		= level,
+		.k		= k,
+		.allocated	= true,
+		/*
+		 * Ensure these keys are done last by journal replay, to unblock
+		 * journal reclaim:
+		 */
+		.journal_seq	= U32_MAX,
+	};
+	struct journal_keys *keys = &c->journal_keys;
+	struct journal_iter *iter;
+	size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
+
+	BUG_ON(test_bit(BCH_FS_RW, &c->flags));
+
+	if (idx < keys->nr &&
+	    journal_key_cmp(&n, &keys->d[idx]) == 0) {
+		if (keys->d[idx].allocated)
+			kfree(keys->d[idx].k);
+		keys->d[idx] = n;
+		return 0;
+	}
+
+	if (keys->nr == keys->size) {
+		struct journal_keys new_keys = {
+			.nr			= keys->nr,
+			.size			= keys->size * 2,
+			.journal_seq_base	= keys->journal_seq_base,
+		};
+
+		new_keys.d = kvmalloc(sizeof(new_keys.d[0]) * new_keys.size, GFP_KERNEL);
+		if (!new_keys.d) {
+			bch_err(c, "%s: error allocating new key array (size %zu)",
+				__func__, new_keys.size);
+			return -ENOMEM;
+		}
+
+		memcpy(new_keys.d, keys->d, sizeof(keys->d[0]) * keys->nr);
+		kvfree(keys->d);
+		*keys = new_keys;
+	}
+
+	array_insert_item(keys->d, keys->nr, idx, n);
+
+	list_for_each_entry(iter, &c->journal_iters, list)
+		journal_iter_fix(c, iter, idx);
+
+	return 0;
+}
+
+/*
+ * Can only be used from the recovery thread while we're still RO - can't be
+ * used once we've got RW, as journal_keys is at that point used by multiple
+ * threads:
+ */
+int bch2_journal_key_insert(struct bch_fs *c, enum btree_id id,
+			    unsigned level, struct bkey_i *k)
+{
+	struct bkey_i *n;
+	int ret;
+
+	n = kmalloc(bkey_bytes(&k->k), GFP_KERNEL);
+	if (!n)
+		return -ENOMEM;
+
+	bkey_copy(n, k);
+	ret = bch2_journal_key_insert_take(c, id, level, n);
+	if (ret)
+		kfree(n);
+	return ret;
+}
+
+int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
+			    unsigned level, struct bpos pos)
+{
+	struct bkey_i whiteout;
+
+	bkey_init(&whiteout.k);
+	whiteout.k.p = pos;
 
-	return l < journal_keys->nr ? journal_keys->d + l : NULL;
+	return bch2_journal_key_insert(c, id, level, &whiteout);
+}
+
+void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
+				  unsigned level, struct bpos pos)
+{
+	struct journal_keys *keys = &c->journal_keys;
+	size_t idx = bch2_journal_key_search(keys, btree, level, pos);
+
+	if (idx < keys->nr &&
+	    keys->d[idx].btree_id	== btree &&
+	    keys->d[idx].level		== level &&
+	    !bpos_cmp(keys->d[idx].k->k.p, pos))
+		keys->d[idx].overwritten = true;
 }
 
 static struct bkey_i *bch2_journal_iter_peek(struct journal_iter *iter)
 {
-	if (iter->k &&
-	    iter->k < iter->keys->d + iter->keys->nr &&
-	    iter->k->btree_id	== iter->btree_id &&
-	    iter->k->level	== iter->level)
-		return iter->k->k;
+	struct journal_key *k = iter->keys->d + iter->idx;
+
+	while (k < iter->keys->d + iter->keys->nr &&
+	       k->btree_id	== iter->btree_id &&
+	       k->level		== iter->level) {
+		if (!k->overwritten)
+			return k->k;
+
+		iter->idx++;
+		k = iter->keys->d + iter->idx;
+	}
 
-	iter->k = NULL;
 	return NULL;
 }
 
 static void bch2_journal_iter_advance(struct journal_iter *iter)
 {
-	if (iter->k)
-		iter->k++;
+	if (iter->idx < iter->keys->nr)
+		iter->idx++;
 }
 
-static void bch2_journal_iter_init(struct journal_iter *iter,
-				   struct journal_keys *journal_keys,
+static void bch2_journal_iter_exit(struct journal_iter *iter)
+{
+	list_del(&iter->list);
+}
+
+static void bch2_journal_iter_init(struct bch_fs *c,
+				   struct journal_iter *iter,
 				   enum btree_id id, unsigned level,
 				   struct bpos pos)
 {
 	iter->btree_id	= id;
 	iter->level	= level;
-	iter->keys	= journal_keys;
-	iter->k		= journal_key_search(journal_keys, id, level, pos);
+	iter->keys	= &c->journal_keys;
+	iter->idx	= bch2_journal_key_search(&c->journal_keys, id, level, pos);
 }
 
 static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
 {
-	return iter->btree
-		? bch2_btree_iter_peek(iter->btree)
-		: bch2_btree_node_iter_peek_unpack(&iter->node_iter,
-						   iter->b, &iter->unpacked);
+	return bch2_btree_node_iter_peek_unpack(&iter->node_iter,
+						iter->b, &iter->unpacked);
 }
 
 static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
 {
-	if (iter->btree)
-		bch2_btree_iter_next(iter->btree);
-	else
-		bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
+	bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
 }
 
 void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
@@ -140,7 +304,7 @@ struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *
 			bkey_i_to_s_c(bch2_journal_iter_peek(&iter->journal));
 
 		if (btree_k.k && journal_k.k) {
-			int cmp = bkey_cmp(btree_k.k->p, journal_k.k->p);
+			int cmp = bpos_cmp(btree_k.k->p, journal_k.k->p);
 
 			if (!cmp)
 				bch2_journal_iter_advance_btree(iter);
@@ -158,8 +322,8 @@ struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *
 		ret = iter->last == journal ? journal_k : btree_k;
 
 		if (iter->b &&
-		    bkey_cmp(ret.k->p, iter->b->data->max_key) > 0) {
-			iter->journal.k = NULL;
+		    bpos_cmp(ret.k->p, iter->b->data->max_key) > 0) {
+			iter->journal.idx = iter->journal.keys->nr;
 			iter->last = none;
 			return bkey_s_c_null;
 		}
@@ -180,99 +344,38 @@ struct bkey_s_c bch2_btree_and_journal_iter_next(struct btree_and_journal_iter *
 	return bch2_btree_and_journal_iter_peek(iter);
 }
 
-void bch2_btree_and_journal_iter_init(struct btree_and_journal_iter *iter,
-				      struct btree_trans *trans,
-				      struct journal_keys *journal_keys,
-				      enum btree_id id, struct bpos pos)
+void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *iter)
 {
-	memset(iter, 0, sizeof(*iter));
-
-	iter->btree = bch2_trans_get_iter(trans, id, pos, BTREE_ITER_PREFETCH);
-	bch2_journal_iter_init(&iter->journal, journal_keys, id, 0, pos);
+	bch2_journal_iter_exit(&iter->journal);
 }
 
-void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
-						struct journal_keys *journal_keys,
-						struct btree *b)
+void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
+						  struct bch_fs *c,
+						  struct btree *b,
+						  struct btree_node_iter node_iter,
+						  struct bpos pos)
 {
 	memset(iter, 0, sizeof(*iter));
 
 	iter->b = b;
-	bch2_btree_node_iter_init_from_start(&iter->node_iter, iter->b);
-	bch2_journal_iter_init(&iter->journal, journal_keys,
-			       b->c.btree_id, b->c.level, b->data->min_key);
-}
-
-/* Walk btree, overlaying keys from the journal: */
-
-static int bch2_btree_and_journal_walk_recurse(struct bch_fs *c, struct btree *b,
-				struct journal_keys *journal_keys,
-				enum btree_id btree_id,
-				btree_walk_node_fn node_fn,
-				btree_walk_key_fn key_fn)
-{
-	struct btree_and_journal_iter iter;
-	struct bkey_s_c k;
-	int ret = 0;
-
-	bch2_btree_and_journal_iter_init_node_iter(&iter, journal_keys, b);
-
-	while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
-		ret = key_fn(c, btree_id, b->c.level, k);
-		if (ret)
-			break;
-
-		if (b->c.level) {
-			struct btree *child;
-			BKEY_PADDED(k) tmp;
-
-			bkey_reassemble(&tmp.k, k);
-			k = bkey_i_to_s_c(&tmp.k);
-
-			bch2_btree_and_journal_iter_advance(&iter);
-
-			if (b->c.level > 0) {
-				child = bch2_btree_node_get_noiter(c, &tmp.k,
-							b->c.btree_id, b->c.level - 1);
-				ret = PTR_ERR_OR_ZERO(child);
-				if (ret)
-					break;
-
-				ret   = (node_fn ? node_fn(c, b) : 0) ?:
-					bch2_btree_and_journal_walk_recurse(c, child,
-						journal_keys, btree_id, node_fn, key_fn);
-				six_unlock_read(&child->c.lock);
-
-				if (ret)
-					break;
-			}
-		} else {
-			bch2_btree_and_journal_iter_advance(&iter);
-		}
-	}
-
-	return ret;
+	iter->node_iter = node_iter;
+	bch2_journal_iter_init(c, &iter->journal, b->c.btree_id, b->c.level, pos);
+	INIT_LIST_HEAD(&iter->journal.list);
 }
 
-int bch2_btree_and_journal_walk(struct bch_fs *c, struct journal_keys *journal_keys,
-				enum btree_id btree_id,
-				btree_walk_node_fn node_fn,
-				btree_walk_key_fn key_fn)
+/*
+ * this version is used by btree_gc before filesystem has gone RW and
+ * multithreaded, so uses the journal_iters list:
+ */
+void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
+						struct bch_fs *c,
+						struct btree *b)
 {
-	struct btree *b = c->btree_roots[btree_id].b;
-	int ret = 0;
-
-	if (btree_node_fake(b))
-		return 0;
-
-	six_lock_read(&b->c.lock, NULL, NULL);
-	ret   = (node_fn ? node_fn(c, b) : 0) ?:
-		bch2_btree_and_journal_walk_recurse(c, b, journal_keys, btree_id,
-						    node_fn, key_fn) ?:
-		key_fn(c, btree_id, b->c.level + 1, bkey_i_to_s_c(&b->key));
-	six_unlock_read(&b->c.lock);
+	struct btree_node_iter node_iter;
 
-	return ret;
+	bch2_btree_node_iter_init_from_start(&node_iter, b);
+	__bch2_btree_and_journal_iter_init_node_iter(iter, c, b, node_iter, b->data->min_key);
+	list_add(&iter->journal.list, &c->journal_iters);
 }
 
 /* sort and dedup all keys in the journal: */
@@ -297,15 +400,19 @@ static int journal_sort_key_cmp(const void *_l, const void *_r)
 	const struct journal_key *l = _l;
 	const struct journal_key *r = _r;
 
-	return  cmp_int(l->btree_id,	r->btree_id) ?:
-		cmp_int(l->level,	r->level) ?:
-		bkey_cmp(l->k->k.p, r->k->k.p) ?:
+	return  journal_key_cmp(l, r) ?:
 		cmp_int(l->journal_seq, r->journal_seq) ?:
 		cmp_int(l->journal_offset, r->journal_offset);
 }
 
 void bch2_journal_keys_free(struct journal_keys *keys)
 {
+	struct journal_key *i;
+
+	for (i = keys->d; i < keys->d + keys->nr; i++)
+		if (i->allocated)
+			kfree(i->k);
+
 	kvfree(keys->d);
 	keys->d = NULL;
 	keys->nr = 0;
@@ -313,7 +420,7 @@ void bch2_journal_keys_free(struct journal_keys *keys)
 
 static struct journal_keys journal_keys_sort(struct list_head *journal_entries)
 {
-	struct journal_replay *p;
+	struct journal_replay *i;
 	struct jset_entry *entry;
 	struct bkey_i *k, *_n;
 	struct journal_keys keys = { NULL };
@@ -323,35 +430,37 @@ static struct journal_keys journal_keys_sort(struct list_head *journal_entries)
 	if (list_empty(journal_entries))
 		return keys;
 
-	keys.journal_seq_base =
-		le64_to_cpu(list_last_entry(journal_entries,
-				struct journal_replay, list)->j.last_seq);
-
-	list_for_each_entry(p, journal_entries, list) {
-		if (le64_to_cpu(p->j.seq) < keys.journal_seq_base)
+	list_for_each_entry(i, journal_entries, list) {
+		if (i->ignore)
 			continue;
 
-		for_each_jset_key(k, _n, entry, &p->j)
+		if (!keys.journal_seq_base)
+			keys.journal_seq_base = le64_to_cpu(i->j.seq);
+
+		for_each_jset_key(k, _n, entry, &i->j)
 			nr_keys++;
 	}
 
+	keys.size = roundup_pow_of_two(nr_keys);
 
-	keys.d = kvmalloc(sizeof(keys.d[0]) * nr_keys, GFP_KERNEL);
+	keys.d = kvmalloc(sizeof(keys.d[0]) * keys.size, GFP_KERNEL);
 	if (!keys.d)
 		goto err;
 
-	list_for_each_entry(p, journal_entries, list) {
-		if (le64_to_cpu(p->j.seq) < keys.journal_seq_base)
+	list_for_each_entry(i, journal_entries, list) {
+		if (i->ignore)
 			continue;
 
-		for_each_jset_key(k, _n, entry, &p->j)
+		BUG_ON(le64_to_cpu(i->j.seq) - keys.journal_seq_base > U32_MAX);
+
+		for_each_jset_key(k, _n, entry, &i->j)
 			keys.d[keys.nr++] = (struct journal_key) {
 				.btree_id	= entry->btree_id,
 				.level		= entry->level,
 				.k		= k,
-				.journal_seq	= le64_to_cpu(p->j.seq) -
+				.journal_seq	= le64_to_cpu(i->j.seq) -
 					keys.journal_seq_base,
-				.journal_offset	= k->_data - p->j._data,
+				.journal_offset	= k->_data - i->j._data,
 			};
 	}
 
@@ -362,7 +471,7 @@ static struct journal_keys journal_keys_sort(struct list_head *journal_entries)
 		while (src + 1 < keys.d + keys.nr &&
 		       src[0].btree_id	== src[1].btree_id &&
 		       src[0].level	== src[1].level &&
-		       !bkey_cmp(src[0].k->k.p, src[1].k->k.p))
+		       !bpos_cmp(src[0].k->k.p, src[1].k->k.p))
 			src++;
 
 		*dst++ = *src++;
@@ -384,252 +493,85 @@ static void replay_now_at(struct journal *j, u64 seq)
 		bch2_journal_pin_put(j, j->replay_journal_seq++);
 }
 
-static int bch2_extent_replay_key(struct bch_fs *c, enum btree_id btree_id,
-				  struct bkey_i *k)
-{
-	struct btree_trans trans;
-	struct btree_iter *iter, *split_iter;
-	/*
-	 * We might cause compressed extents to be split, so we need to pass in
-	 * a disk_reservation:
-	 */
-	struct disk_reservation disk_res =
-		bch2_disk_reservation_init(c, 0);
-	struct bkey_i *split;
-	struct bpos atomic_end;
-	/*
-	 * Some extents aren't equivalent - w.r.t. what the triggers do
-	 * - if they're split:
-	 */
-	bool remark_if_split = bch2_bkey_sectors_compressed(bkey_i_to_s_c(k)) ||
-		k->k.type == KEY_TYPE_reflink_p;
-	bool remark = false;
-	int ret;
-
-	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
-retry:
-	bch2_trans_begin(&trans);
-
-	iter = bch2_trans_get_iter(&trans, btree_id,
-				   bkey_start_pos(&k->k),
-				   BTREE_ITER_INTENT);
-
-	do {
-		ret = bch2_btree_iter_traverse(iter);
-		if (ret)
-			goto err;
-
-		atomic_end = bpos_min(k->k.p, iter->l[0].b->key.k.p);
-
-		split = bch2_trans_kmalloc(&trans, bkey_bytes(&k->k));
-		ret = PTR_ERR_OR_ZERO(split);
-		if (ret)
-			goto err;
-
-		if (!remark &&
-		    remark_if_split &&
-		    bkey_cmp(atomic_end, k->k.p) < 0) {
-			ret = bch2_disk_reservation_add(c, &disk_res,
-					k->k.size *
-					bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(k)),
-					BCH_DISK_RESERVATION_NOFAIL);
-			BUG_ON(ret);
-
-			remark = true;
-		}
-
-		bkey_copy(split, k);
-		bch2_cut_front(iter->pos, split);
-		bch2_cut_back(atomic_end, split);
-
-		split_iter = bch2_trans_copy_iter(&trans, iter);
-
-		/*
-		 * It's important that we don't go through the
-		 * extent_handle_overwrites() and extent_update_to_keys() path
-		 * here: journal replay is supposed to treat extents like
-		 * regular keys
-		 */
-		__bch2_btree_iter_set_pos(split_iter, split->k.p, false);
-		bch2_trans_update(&trans, split_iter, split,
-				  BTREE_TRIGGER_NORUN);
-		bch2_trans_iter_put(&trans, split_iter);
-
-		bch2_btree_iter_set_pos(iter, split->k.p);
-
-		if (remark) {
-			ret = bch2_trans_mark_key(&trans, bkey_i_to_s_c(split),
-						  0, split->k.size,
-						  BTREE_TRIGGER_INSERT);
-			if (ret)
-				goto err;
-		}
-	} while (bkey_cmp(iter->pos, k->k.p) < 0);
-
-	if (remark) {
-		ret = bch2_trans_mark_key(&trans, bkey_i_to_s_c(k),
-					  0, -((s64) k->k.size),
-					  BTREE_TRIGGER_OVERWRITE);
-		if (ret)
-			goto err;
-	}
-
-	ret = bch2_trans_commit(&trans, &disk_res, NULL,
-				BTREE_INSERT_NOFAIL|
-				BTREE_INSERT_LAZY_RW|
-				BTREE_INSERT_JOURNAL_REPLAY);
-err:
-	bch2_trans_iter_put(&trans, iter);
-
-	if (ret == -EINTR)
-		goto retry;
-
-	bch2_disk_reservation_put(c, &disk_res);
-
-	return bch2_trans_exit(&trans) ?: ret;
-}
-
-static int __bch2_journal_replay_key(struct btree_trans *trans,
-				     enum btree_id id, unsigned level,
-				     struct bkey_i *k)
+static int bch2_journal_replay_key(struct btree_trans *trans,
+				   struct journal_key *k)
 {
-	struct btree_iter *iter;
+	struct btree_iter iter;
+	unsigned iter_flags =
+		BTREE_ITER_INTENT|
+		BTREE_ITER_NOT_EXTENTS;
 	int ret;
 
-	iter = bch2_trans_get_node_iter(trans, id, k->k.p,
-					BTREE_MAX_DEPTH, level,
-					BTREE_ITER_INTENT);
-
-	/*
-	 * iter->flags & BTREE_ITER_IS_EXTENTS triggers the update path to run
-	 * extent_handle_overwrites() and extent_update_to_keys() - but we don't
-	 * want that here, journal replay is supposed to treat extents like
-	 * regular keys:
-	 */
-	__bch2_btree_iter_set_pos(iter, k->k.p, false);
-
-	ret   = bch2_btree_iter_traverse(iter) ?:
-		bch2_trans_update(trans, iter, k, BTREE_TRIGGER_NORUN);
-	bch2_trans_iter_put(trans, iter);
-	return ret;
-}
+	if (!k->level && k->btree_id == BTREE_ID_alloc)
+		iter_flags |= BTREE_ITER_CACHED;
 
-static int bch2_journal_replay_key(struct bch_fs *c, enum btree_id id,
-				   unsigned level, struct bkey_i *k)
-{
-	return bch2_trans_do(c, NULL, NULL,
-			     BTREE_INSERT_NOFAIL|
-			     BTREE_INSERT_LAZY_RW|
-			     BTREE_INSERT_JOURNAL_REPLAY,
-			     __bch2_journal_replay_key(&trans, id, level, k));
-}
+	bch2_trans_node_iter_init(trans, &iter, k->btree_id, k->k->k.p,
+				  BTREE_MAX_DEPTH, k->level,
+				  iter_flags);
+	ret = bch2_btree_iter_traverse(&iter);
+	if (ret)
+		goto out;
 
-static int __bch2_alloc_replay_key(struct btree_trans *trans, struct bkey_i *k)
-{
-	struct btree_iter *iter;
-	int ret;
+	/* Must be checked with btree locked: */
+	if (k->overwritten)
+		goto out;
 
-	iter = bch2_trans_get_iter(trans, BTREE_ID_ALLOC, k->k.p,
-				   BTREE_ITER_CACHED|
-				   BTREE_ITER_CACHED_NOFILL|
-				   BTREE_ITER_INTENT);
-	ret = bch2_trans_update(trans, iter, k, BTREE_TRIGGER_NORUN);
-	bch2_trans_iter_put(trans, iter);
+	ret = bch2_trans_update(trans, &iter, k->k, BTREE_TRIGGER_NORUN);
+out:
+	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
 
-static int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k)
-{
-	return bch2_trans_do(c, NULL, NULL,
-			     BTREE_INSERT_NOFAIL|
-			     BTREE_INSERT_USE_RESERVE|
-			     BTREE_INSERT_LAZY_RW|
-			     BTREE_INSERT_JOURNAL_REPLAY,
-			__bch2_alloc_replay_key(&trans, k));
-}
-
 static int journal_sort_seq_cmp(const void *_l, const void *_r)
 {
-	const struct journal_key *l = _l;
-	const struct journal_key *r = _r;
+	const struct journal_key *l = *((const struct journal_key **)_l);
+	const struct journal_key *r = *((const struct journal_key **)_r);
 
-	return  cmp_int(r->level,	l->level) ?:
-		cmp_int(l->journal_seq, r->journal_seq) ?:
-		cmp_int(l->btree_id,	r->btree_id) ?:
-		bkey_cmp(l->k->k.p,	r->k->k.p);
+	return cmp_int(l->journal_seq, r->journal_seq);
 }
 
-static int bch2_journal_replay(struct bch_fs *c,
-			       struct journal_keys keys)
+static int bch2_journal_replay(struct bch_fs *c)
 {
+	struct journal_keys *keys = &c->journal_keys;
+	struct journal_key **keys_sorted, *k;
 	struct journal *j = &c->journal;
-	struct journal_key *i;
-	u64 seq;
+	size_t i;
 	int ret;
 
-	sort(keys.d, keys.nr, sizeof(keys.d[0]), journal_sort_seq_cmp, NULL);
+	keys_sorted = kvmalloc_array(sizeof(*keys_sorted), keys->nr, GFP_KERNEL);
+	if (!keys_sorted)
+		return -ENOMEM;
 
-	if (keys.nr)
-		replay_now_at(j, keys.journal_seq_base);
-
-	seq = j->replay_journal_seq;
-
-	/*
-	 * First replay updates to the alloc btree - these will only update the
-	 * btree key cache:
-	 */
-	for_each_journal_key(keys, i) {
-		cond_resched();
-
-		if (!i->level && i->btree_id == BTREE_ID_ALLOC) {
-			j->replay_journal_seq = keys.journal_seq_base + i->journal_seq;
-			ret = bch2_alloc_replay_key(c, i->k);
-			if (ret)
-				goto err;
-		}
-	}
-
-	/*
-	 * Next replay updates to interior btree nodes:
-	 */
-	for_each_journal_key(keys, i) {
-		cond_resched();
+	for (i = 0; i < keys->nr; i++)
+		keys_sorted[i] = &keys->d[i];
 
-		if (i->level) {
-			j->replay_journal_seq = keys.journal_seq_base + i->journal_seq;
-			ret = bch2_journal_replay_key(c, i->btree_id, i->level, i->k);
-			if (ret)
-				goto err;
-		}
-	}
+	sort(keys_sorted, keys->nr,
+	     sizeof(keys_sorted[0]),
+	     journal_sort_seq_cmp, NULL);
 
-	/*
-	 * Now that the btree is in a consistent state, we can start journal
-	 * reclaim (which will be flushing entries from the btree key cache back
-	 * to the btree:
-	 */
-	set_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags);
-	set_bit(JOURNAL_RECLAIM_STARTED, &j->flags);
-	journal_reclaim_kick(j);
+	if (keys->nr)
+		replay_now_at(j, keys->journal_seq_base);
 
-	j->replay_journal_seq = seq;
+	for (i = 0; i < keys->nr; i++) {
+		k = keys_sorted[i];
 
-	/*
-	 * Now replay leaf node updates:
-	 */
-	for_each_journal_key(keys, i) {
 		cond_resched();
 
-		if (i->level || i->btree_id == BTREE_ID_ALLOC)
-			continue;
-
-		replay_now_at(j, keys.journal_seq_base + i->journal_seq);
+		if (!k->allocated)
+			replay_now_at(j, keys->journal_seq_base + k->journal_seq);
 
-		ret = i->k->k.size
-			? bch2_extent_replay_key(c, i->btree_id, i->k)
-			: bch2_journal_replay_key(c, i->btree_id, i->level, i->k);
-		if (ret)
+		ret = bch2_trans_do(c, NULL, NULL,
+				    BTREE_INSERT_LAZY_RW|
+				    BTREE_INSERT_NOFAIL|
+				    (!k->allocated
+				     ? BTREE_INSERT_JOURNAL_REPLAY|JOURNAL_WATERMARK_reserved
+				     : 0),
+			     bch2_journal_replay_key(&trans, k));
+		if (ret) {
+			bch_err(c, "journal replay: error %d while replaying key at btree %s level %u",
+				ret, bch2_btree_ids[k->btree_id], k->level);
 			goto err;
+		}
 	}
 
 	replay_now_at(j, j->replay_journal_seq_end);
@@ -637,49 +579,12 @@ static int bch2_journal_replay(struct bch_fs *c,
 
 	bch2_journal_set_replay_done(j);
 	bch2_journal_flush_all_pins(j);
-	return bch2_journal_error(j);
-err:
-	bch_err(c, "journal replay: error %d while replaying key", ret);
-	return ret;
-}
-
-static bool journal_empty(struct list_head *journal)
-{
-	return list_empty(journal) ||
-		journal_entry_empty(&list_last_entry(journal,
-					struct journal_replay, list)->j);
-}
-
-static int
-verify_journal_entries_not_blacklisted_or_missing(struct bch_fs *c,
-						  struct list_head *journal)
-{
-	struct journal_replay *i =
-		list_last_entry(journal, struct journal_replay, list);
-	u64 start_seq	= le64_to_cpu(i->j.last_seq);
-	u64 end_seq	= le64_to_cpu(i->j.seq);
-	u64 seq		= start_seq;
-	int ret = 0;
+	ret = bch2_journal_error(j);
 
-	list_for_each_entry(i, journal, list) {
-		if (le64_to_cpu(i->j.seq) < start_seq)
-			continue;
-
-		fsck_err_on(seq != le64_to_cpu(i->j.seq), c,
-			"journal entries %llu-%llu missing! (replaying %llu-%llu)",
-			seq, le64_to_cpu(i->j.seq) - 1,
-			start_seq, end_seq);
-
-		seq = le64_to_cpu(i->j.seq);
-
-		fsck_err_on(bch2_journal_seq_is_blacklisted(c, seq, false), c,
-			    "found blacklisted journal entry %llu", seq);
-
-		do {
-			seq++;
-		} while (bch2_journal_seq_is_blacklisted(c, seq, false));
-	}
-fsck_err:
+	if (keys->nr && !ret)
+		bch2_journal_log_msg(&c->journal, "journal replay finished");
+err:
+	kvfree(keys_sorted);
 	return ret;
 }
 
@@ -717,15 +622,15 @@ static int journal_replay_entry_early(struct bch_fs *c,
 			container_of(entry, struct jset_entry_usage, entry);
 
 		switch (entry->btree_id) {
-		case FS_USAGE_RESERVED:
+		case BCH_FS_USAGE_reserved:
 			if (entry->level < BCH_REPLICAS_MAX)
 				c->usage_base->persistent_reserved[entry->level] =
 					le64_to_cpu(u->v);
 			break;
-		case FS_USAGE_INODES:
+		case BCH_FS_USAGE_inodes:
 			c->usage_base->nr_inodes = le64_to_cpu(u->v);
 			break;
-		case FS_USAGE_KEY_VERSION:
+		case BCH_FS_USAGE_key_version:
 			atomic64_set(&c->key_version,
 				     le64_to_cpu(u->v));
 			break;
@@ -736,10 +641,28 @@ static int journal_replay_entry_early(struct bch_fs *c,
 	case BCH_JSET_ENTRY_data_usage: {
 		struct jset_entry_data_usage *u =
 			container_of(entry, struct jset_entry_data_usage, entry);
+
 		ret = bch2_replicas_set_usage(c, &u->r,
 					      le64_to_cpu(u->v));
 		break;
 	}
+	case BCH_JSET_ENTRY_dev_usage: {
+		struct jset_entry_dev_usage *u =
+			container_of(entry, struct jset_entry_dev_usage, entry);
+		struct bch_dev *ca = bch_dev_bkey_exists(c, le32_to_cpu(u->dev));
+		unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
+
+		ca->usage_base->buckets_ec		= le64_to_cpu(u->buckets_ec);
+		ca->usage_base->buckets_unavailable	= le64_to_cpu(u->buckets_unavailable);
+
+		for (i = 0; i < min_t(unsigned, nr_types, BCH_DATA_NR); i++) {
+			ca->usage_base->d[i].buckets	= le64_to_cpu(u->d[i].buckets);
+			ca->usage_base->d[i].sectors	= le64_to_cpu(u->d[i].sectors);
+			ca->usage_base->d[i].fragmented	= le64_to_cpu(u->d[i].fragmented);
+		}
+
+		break;
+	}
 	case BCH_JSET_ENTRY_blacklist: {
 		struct jset_entry_blacklist *bl_entry =
 			container_of(entry, struct jset_entry_blacklist, entry);
@@ -758,6 +681,12 @@ static int journal_replay_entry_early(struct bch_fs *c,
 				le64_to_cpu(bl_entry->end) + 1);
 		break;
 	}
+	case BCH_JSET_ENTRY_clock: {
+		struct jset_entry_clock *clock =
+			container_of(entry, struct jset_entry_clock, entry);
+
+		atomic64_set(&c->io_clock[clock->rw].now, le64_to_cpu(clock->time));
+	}
 	}
 
 	return ret;
@@ -767,13 +696,11 @@ static int journal_replay_early(struct bch_fs *c,
 				struct bch_sb_field_clean *clean,
 				struct list_head *journal)
 {
+	struct journal_replay *i;
 	struct jset_entry *entry;
 	int ret;
 
 	if (clean) {
-		c->bucket_clock[READ].hand = le16_to_cpu(clean->read_clock);
-		c->bucket_clock[WRITE].hand = le16_to_cpu(clean->write_clock);
-
 		for (entry = clean->start;
 		     entry != vstruct_end(&clean->field);
 		     entry = vstruct_next(entry)) {
@@ -782,18 +709,16 @@ static int journal_replay_early(struct bch_fs *c,
 				return ret;
 		}
 	} else {
-		struct journal_replay *i =
-			list_last_entry(journal, struct journal_replay, list);
+		list_for_each_entry(i, journal, list) {
+			if (i->ignore)
+				continue;
 
-		c->bucket_clock[READ].hand = le16_to_cpu(i->j.read_clock);
-		c->bucket_clock[WRITE].hand = le16_to_cpu(i->j.write_clock);
-
-		list_for_each_entry(i, journal, list)
 			vstruct_for_each(&i->j, entry) {
 				ret = journal_replay_entry_early(c, entry);
 				if (ret)
 					return ret;
 			}
+		}
 	}
 
 	bch2_fs_usage_initialize(c);
@@ -840,11 +765,10 @@ static int verify_superblock_clean(struct bch_fs *c,
 {
 	unsigned i;
 	struct bch_sb_field_clean *clean = *cleanp;
+	struct printbuf buf1 = PRINTBUF;
+	struct printbuf buf2 = PRINTBUF;
 	int ret = 0;
 
-	if (!c->sb.clean || !j)
-		return 0;
-
 	if (mustfix_fsck_err_on(j->seq != clean->journal_seq, c,
 			"superblock journal seq (%llu) doesn't match journal (%llu) after clean shutdown",
 			le64_to_cpu(clean->journal_seq),
@@ -854,15 +778,7 @@ static int verify_superblock_clean(struct bch_fs *c,
 		return 0;
 	}
 
-	mustfix_fsck_err_on(j->read_clock != clean->read_clock, c,
-			"superblock read clock %u doesn't match journal %u after clean shutdown",
-			clean->read_clock, j->read_clock);
-	mustfix_fsck_err_on(j->write_clock != clean->write_clock, c,
-			"superblock write clock %u doesn't match journal %u after clean shutdown",
-			clean->write_clock, j->write_clock);
-
 	for (i = 0; i < BTREE_ID_NR; i++) {
-		char buf1[200], buf2[200];
 		struct bkey_i *k1, *k2;
 		unsigned l1 = 0, l2 = 0;
 
@@ -872,6 +788,19 @@ static int verify_superblock_clean(struct bch_fs *c,
 		if (!k1 && !k2)
 			continue;
 
+		printbuf_reset(&buf1);
+		printbuf_reset(&buf2);
+
+		if (k1)
+			bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(k1));
+		else
+			pr_buf(&buf1, "(none)");
+
+		if (k2)
+			bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(k2));
+		else
+			pr_buf(&buf2, "(none)");
+
 		mustfix_fsck_err_on(!k1 || !k2 ||
 				    IS_ERR(k1) ||
 				    IS_ERR(k2) ||
@@ -881,10 +810,12 @@ static int verify_superblock_clean(struct bch_fs *c,
 			"superblock btree root %u doesn't match journal after clean shutdown\n"
 			"sb:      l=%u %s\n"
 			"journal: l=%u %s\n", i,
-			l1, (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(k1)), buf1),
-			l2, (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(k2)), buf2));
+			l1, buf1.buf,
+			l2, buf2.buf);
 	}
 fsck_err:
+	printbuf_exit(&buf2);
+	printbuf_exit(&buf1);
 	return ret;
 }
 
@@ -911,9 +842,11 @@ static struct bch_sb_field_clean *read_superblock_clean(struct bch_fs *c)
 		return ERR_PTR(-ENOMEM);
 	}
 
-	if (le16_to_cpu(c->disk_sb.sb->version) <
-	    bcachefs_metadata_version_bkey_renumber)
-		bch2_sb_clean_renumber(clean, READ);
+	ret = bch2_sb_clean_validate_late(c, clean, READ);
+	if (ret) {
+		mutex_unlock(&c->sb_lock);
+		return ERR_PTR(ret);
+	}
 
 	mutex_unlock(&c->sb_lock);
 
@@ -934,29 +867,29 @@ static int read_btree_roots(struct bch_fs *c)
 		if (!r->alive)
 			continue;
 
-		if (i == BTREE_ID_ALLOC &&
+		if (i == BTREE_ID_alloc &&
 		    c->opts.reconstruct_alloc) {
-			c->sb.compat &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_INFO);
+			c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
 			continue;
 		}
 
 		if (r->error) {
-			__fsck_err(c, i == BTREE_ID_ALLOC
+			__fsck_err(c, i == BTREE_ID_alloc
 				   ? FSCK_CAN_IGNORE : 0,
 				   "invalid btree root %s",
 				   bch2_btree_ids[i]);
-			if (i == BTREE_ID_ALLOC)
-				c->sb.compat &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_INFO);
+			if (i == BTREE_ID_alloc)
+				c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
 		}
 
 		ret = bch2_btree_root_read(c, i, &r->key, r->level);
 		if (ret) {
-			__fsck_err(c, i == BTREE_ID_ALLOC
+			__fsck_err(c, i == BTREE_ID_alloc
 				   ? FSCK_CAN_IGNORE : 0,
 				   "error reading btree root %s",
 				   bch2_btree_ids[i]);
-			if (i == BTREE_ID_ALLOC)
-				c->sb.compat &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_INFO);
+			if (i == BTREE_ID_alloc)
+				c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
 		}
 	}
 
@@ -967,13 +900,81 @@ fsck_err:
 	return ret;
 }
 
+static int bch2_fs_initialize_subvolumes(struct bch_fs *c)
+{
+	struct bkey_i_snapshot	root_snapshot;
+	struct bkey_i_subvolume root_volume;
+	int ret;
+
+	bkey_snapshot_init(&root_snapshot.k_i);
+	root_snapshot.k.p.offset = U32_MAX;
+	root_snapshot.v.flags	= 0;
+	root_snapshot.v.parent	= 0;
+	root_snapshot.v.subvol	= BCACHEFS_ROOT_SUBVOL;
+	root_snapshot.v.pad	= 0;
+	SET_BCH_SNAPSHOT_SUBVOL(&root_snapshot.v, true);
+
+	ret = bch2_btree_insert(c, BTREE_ID_snapshots,
+				&root_snapshot.k_i,
+				NULL, NULL, 0);
+	if (ret)
+		return ret;
+
+
+	bkey_subvolume_init(&root_volume.k_i);
+	root_volume.k.p.offset = BCACHEFS_ROOT_SUBVOL;
+	root_volume.v.flags	= 0;
+	root_volume.v.snapshot	= cpu_to_le32(U32_MAX);
+	root_volume.v.inode	= cpu_to_le64(BCACHEFS_ROOT_INO);
+
+	ret = bch2_btree_insert(c, BTREE_ID_subvolumes,
+				&root_volume.k_i,
+				NULL, NULL, 0);
+	if (ret)
+		return ret;
+
+	return 0;
+}
+
+static int bch2_fs_upgrade_for_subvolumes(struct btree_trans *trans)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	struct bch_inode_unpacked inode;
+	int ret;
+
+	bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes,
+			     SPOS(0, BCACHEFS_ROOT_INO, U32_MAX), 0);
+	k = bch2_btree_iter_peek_slot(&iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	if (!bkey_is_inode(k.k)) {
+		bch_err(trans->c, "root inode not found");
+		ret = -ENOENT;
+		goto err;
+	}
+
+	ret = bch2_inode_unpack(k, &inode);
+	BUG_ON(ret);
+
+	inode.bi_subvol = BCACHEFS_ROOT_SUBVOL;
+
+	ret = bch2_inode_write(trans, &iter, &inode);
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
 int bch2_fs_recovery(struct bch_fs *c)
 {
 	const char *err = "cannot allocate memory";
 	struct bch_sb_field_clean *clean = NULL;
-	u64 journal_seq;
-	bool write_sb = false, need_write_alloc = false;
-	int ret;
+	struct jset *last_journal_entry = NULL;
+	u64 blacklist_seq, journal_seq;
+	bool write_sb = false;
+	int ret = 0;
 
 	if (c->sb.clean)
 		clean = read_superblock_clean(c);
@@ -984,6 +985,33 @@ int bch2_fs_recovery(struct bch_fs *c)
 	if (c->sb.clean)
 		bch_info(c, "recovering from clean shutdown, journal seq %llu",
 			 le64_to_cpu(clean->journal_seq));
+	else
+		bch_info(c, "recovering from unclean shutdown");
+
+	if (!(c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))) {
+		bch_err(c, "feature new_extent_overwrite not set, filesystem no longer supported");
+		ret = -EINVAL;
+		goto err;
+	}
+
+	if (!c->sb.clean &&
+	    !(c->sb.features & (1ULL << BCH_FEATURE_extents_above_btree_updates))) {
+		bch_err(c, "filesystem needs recovery from older version; run fsck from older bcachefs-tools to fix");
+		ret = -EINVAL;
+		goto err;
+	}
+
+	if (!(c->sb.compat & (1ULL << BCH_COMPAT_bformat_overflow_done))) {
+		bch_err(c, "filesystem may have incompatible bkey formats; run fsck from the compat branch to fix");
+		ret = -EINVAL;
+		goto err;
+	}
+
+	if (!(c->sb.features & (1ULL << BCH_FEATURE_alloc_v2))) {
+		bch_info(c, "alloc_v2 feature bit not set, fsck required");
+		c->opts.fsck = true;
+		c->opts.fix_errors = FSCK_OPT_YES;
+	}
 
 	if (!c->replicas.entries ||
 	    c->opts.rebuild_replicas) {
@@ -991,24 +1019,55 @@ int bch2_fs_recovery(struct bch_fs *c)
 		set_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
 	}
 
+	if (!c->opts.nochanges) {
+		if (c->sb.version < bcachefs_metadata_version_inode_backpointers) {
+			bch_info(c, "version prior to inode backpointers, upgrade and fsck required");
+			c->opts.version_upgrade	= true;
+			c->opts.fsck		= true;
+			c->opts.fix_errors	= FSCK_OPT_YES;
+		} else if (c->sb.version < bcachefs_metadata_version_subvol_dirent) {
+			bch_info(c, "filesystem version is prior to subvol_dirent - upgrading");
+			c->opts.version_upgrade = true;
+			c->opts.fsck		= true;
+		} else if (c->sb.version < bcachefs_metadata_version_alloc_v4) {
+			bch_info(c, "filesystem version is prior to alloc_v4 - upgrading");
+			c->opts.version_upgrade = true;
+		}
+	}
+
+	ret = bch2_blacklist_table_initialize(c);
+	if (ret) {
+		bch_err(c, "error initializing blacklist table");
+		goto err;
+	}
+
 	if (!c->sb.clean || c->opts.fsck || c->opts.keep_journal) {
-		struct jset *j;
+		struct journal_replay *i;
 
-		ret = bch2_journal_read(c, &c->journal_entries);
+		bch_verbose(c, "starting journal read");
+		ret = bch2_journal_read(c, &c->journal_entries,
+					&blacklist_seq, &journal_seq);
 		if (ret)
 			goto err;
 
-		if (mustfix_fsck_err_on(c->sb.clean && !journal_empty(&c->journal_entries), c,
+		list_for_each_entry_reverse(i, &c->journal_entries, list)
+			if (!i->ignore) {
+				last_journal_entry = &i->j;
+				break;
+			}
+
+		if (mustfix_fsck_err_on(c->sb.clean &&
+					last_journal_entry &&
+					!journal_entry_empty(last_journal_entry), c,
 				"filesystem marked clean but journal not empty")) {
-			c->sb.compat &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_INFO);
+			c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
 			SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
 			c->sb.clean = false;
 		}
 
-		if (!c->sb.clean && list_empty(&c->journal_entries)) {
-			bch_err(c, "no journal entries found");
-			ret = BCH_FSCK_REPAIR_IMPOSSIBLE;
-			goto err;
+		if (!last_journal_entry) {
+			fsck_err_on(!c->sb.clean, c, "no journal entries found");
+			goto use_clean;
 		}
 
 		c->journal_keys = journal_keys_sort(&c->journal_entries);
@@ -1017,58 +1076,54 @@ int bch2_fs_recovery(struct bch_fs *c)
 			goto err;
 		}
 
-		j = &list_last_entry(&c->journal_entries,
-				     struct journal_replay, list)->j;
-
-		ret = verify_superblock_clean(c, &clean, j);
-		if (ret)
+		if (c->sb.clean && last_journal_entry) {
+			ret = verify_superblock_clean(c, &clean,
+						      last_journal_entry);
+			if (ret)
+				goto err;
+		}
+	} else {
+use_clean:
+		if (!clean) {
+			bch_err(c, "no superblock clean section found");
+			ret = BCH_FSCK_REPAIR_IMPOSSIBLE;
 			goto err;
 
-		journal_seq = le64_to_cpu(j->seq) + 1;
-	} else {
-		journal_seq = le64_to_cpu(clean->journal_seq) + 1;
+		}
+		blacklist_seq = journal_seq = le64_to_cpu(clean->journal_seq) + 1;
 	}
 
-	if (!c->sb.clean &&
-	    !(c->sb.features & (1ULL << BCH_FEATURE_extents_above_btree_updates))) {
-		bch_err(c, "filesystem needs recovery from older version; run fsck from older bcachefs-tools to fix");
-		ret = -EINVAL;
-		goto err;
-	}
+	if (c->opts.read_journal_only)
+		goto out;
 
 	if (c->opts.reconstruct_alloc) {
-		c->sb.compat &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_INFO);
+		c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
 		drop_alloc_keys(&c->journal_keys);
 	}
 
+	zero_out_btree_mem_ptr(&c->journal_keys);
+
 	ret = journal_replay_early(c, clean, &c->journal_entries);
 	if (ret)
 		goto err;
 
-	if (!c->sb.clean) {
+	/*
+	 * After an unclean shutdown, skip then next few journal sequence
+	 * numbers as they may have been referenced by btree writes that
+	 * happened before their corresponding journal writes - those btree
+	 * writes need to be ignored, by skipping and blacklisting the next few
+	 * journal sequence numbers:
+	 */
+	if (!c->sb.clean)
+		journal_seq += 8;
+
+	if (blacklist_seq != journal_seq) {
 		ret = bch2_journal_seq_blacklist_add(c,
-						     journal_seq,
-						     journal_seq + 4);
+					blacklist_seq, journal_seq);
 		if (ret) {
 			bch_err(c, "error creating new journal seq blacklist entry");
 			goto err;
 		}
-
-		journal_seq += 4;
-
-		/*
-		 * The superblock needs to be written before we do any btree
-		 * node writes: it will be in the read_write() path
-		 */
-	}
-
-	ret = bch2_blacklist_table_initialize(c);
-
-	if (!list_empty(&c->journal_entries)) {
-		ret = verify_journal_entries_not_blacklisted_or_missing(c,
-							&c->journal_entries);
-		if (ret)
-			goto err;
 	}
 
 	ret = bch2_fs_journal_start(&c->journal, journal_seq,
@@ -1082,52 +1137,61 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 	bch_verbose(c, "starting alloc read");
 	err = "error reading allocation information";
-	ret = bch2_alloc_read(c, &c->journal_keys);
+
+	down_read(&c->gc_lock);
+	ret = bch2_alloc_read(c);
+	up_read(&c->gc_lock);
+
 	if (ret)
 		goto err;
 	bch_verbose(c, "alloc read done");
 
 	bch_verbose(c, "starting stripes_read");
 	err = "error reading stripes";
-	ret = bch2_stripes_read(c, &c->journal_keys);
+	ret = bch2_stripes_read(c);
 	if (ret)
 		goto err;
 	bch_verbose(c, "stripes_read done");
 
-	set_bit(BCH_FS_ALLOC_READ_DONE, &c->flags);
+	/*
+	 * If we're not running fsck, this ensures bch2_fsck_err() calls are
+	 * instead interpreted as bch2_inconsistent_err() calls:
+	 */
+	if (!c->opts.fsck)
+		set_bit(BCH_FS_FSCK_DONE, &c->flags);
 
-	if ((c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_INFO)) &&
-	    !(c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_METADATA))) {
-		/*
-		 * interior btree node updates aren't consistent with the
-		 * journal; after an unclean shutdown we have to walk all
-		 * pointers to metadata:
-		 */
-		bch_info(c, "starting metadata mark and sweep");
-		err = "error in mark and sweep";
-		ret = bch2_gc(c, &c->journal_keys, true, true);
-		if (ret < 0)
-			goto err;
+	if (c->opts.fsck ||
+	    !(c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)) ||
+	    !(c->sb.compat & (1ULL << BCH_COMPAT_alloc_metadata)) ||
+	    test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags)) {
+		bool metadata_only = c->opts.norecovery;
+
+		bch_info(c, "checking allocations");
+		err = "error checking allocations";
+		ret = bch2_gc(c, true, metadata_only);
 		if (ret)
-			need_write_alloc = true;
-		bch_verbose(c, "mark and sweep done");
+			goto err;
+		bch_verbose(c, "done checking allocations");
 	}
 
-	if (c->opts.fsck ||
-	    !(c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_INFO)) ||
-	    test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags)) {
-		bch_info(c, "starting mark and sweep");
-		err = "error in mark and sweep";
-		ret = bch2_gc(c, &c->journal_keys, true, false);
-		if (ret < 0)
+	if (c->opts.fsck) {
+		bch_info(c, "checking need_discard and freespace btrees");
+		err = "error checking need_discard and freespace btrees";
+		ret = bch2_check_alloc_info(c, true);
+		if (ret)
 			goto err;
+
+		ret = bch2_check_lrus(c, true);
 		if (ret)
-			need_write_alloc = true;
-		bch_verbose(c, "mark and sweep done");
+			goto err;
+		bch_verbose(c, "done checking need_discard and freespace btrees");
 	}
 
+	bch2_stripes_heap_start(c);
+
 	clear_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
 	set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
+	set_bit(BCH_FS_MAY_GO_RW, &c->flags);
 
 	/*
 	 * Skip past versions that might have possibly been used (as nonces),
@@ -1139,49 +1203,42 @@ int bch2_fs_recovery(struct bch_fs *c)
 	if (c->opts.norecovery)
 		goto out;
 
-	bch_verbose(c, "starting journal replay");
+	bch_verbose(c, "starting journal replay, %zu keys", c->journal_keys.nr);
 	err = "journal replay failed";
-	ret = bch2_journal_replay(c, c->journal_keys);
+	ret = bch2_journal_replay(c);
 	if (ret)
 		goto err;
-	bch_verbose(c, "journal replay done");
+	if (c->opts.verbose || !c->sb.clean)
+		bch_info(c, "journal replay done");
 
-	if (need_write_alloc && !c->opts.nochanges) {
-		/*
-		 * note that even when filesystem was clean there might be work
-		 * to do here, if we ran gc (because of fsck) which recalculated
-		 * oldest_gen:
-		 */
-		bch_verbose(c, "writing allocation info");
-		err = "error writing out alloc info";
-		ret = bch2_stripes_write(c, BTREE_INSERT_LAZY_RW) ?:
-			bch2_alloc_write(c, BTREE_INSERT_LAZY_RW);
-		if (ret) {
-			bch_err(c, "error writing alloc info");
-			goto err;
-		}
-		bch_verbose(c, "alloc write done");
+	err = "error initializing freespace";
+	ret = bch2_fs_freespace_init(c);
+	if (ret)
+		goto err;
+
+	if (c->sb.version < bcachefs_metadata_version_snapshot_2) {
+		bch2_fs_lazy_rw(c);
 
-		set_bit(BCH_FS_ALLOC_WRITTEN, &c->flags);
+		err = "error creating root snapshot node";
+		ret = bch2_fs_initialize_subvolumes(c);
+		if (ret)
+			goto err;
 	}
 
-	if (!c->sb.clean) {
-		if (!(c->sb.features & (1 << BCH_FEATURE_atomic_nlink))) {
-			bch_info(c, "checking inode link counts");
-			err = "error in recovery";
-			ret = bch2_fsck_inode_nlink(c);
-			if (ret)
-				goto err;
-			bch_verbose(c, "check inodes done");
+	bch_verbose(c, "reading snapshots table");
+	err = "error reading snapshots table";
+	ret = bch2_fs_snapshots_start(c);
+	if (ret)
+		goto err;
+	bch_verbose(c, "reading snapshots done");
 
-		} else {
-			bch_verbose(c, "checking for deleted inodes");
-			err = "error in recovery";
-			ret = bch2_fsck_walk_inodes_only(c);
-			if (ret)
-				goto err;
-			bch_verbose(c, "check inodes done");
-		}
+	if (c->sb.version < bcachefs_metadata_version_snapshot_2) {
+		/* set bi_subvol on root inode */
+		err = "error upgrade root inode for subvolumes";
+		ret = bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
+				    bch2_fs_upgrade_for_subvolumes(&trans));
+		if (ret)
+			goto err;
 	}
 
 	if (c->opts.fsck) {
@@ -1191,6 +1248,13 @@ int bch2_fs_recovery(struct bch_fs *c)
 		if (ret)
 			goto err;
 		bch_verbose(c, "fsck done");
+	} else if (!c->sb.clean) {
+		bch_verbose(c, "checking for deleted inodes");
+		err = "error in recovery";
+		ret = bch2_fsck_walk_inodes_only(c);
+		if (ret)
+			goto err;
+		bch_verbose(c, "check inodes done");
 	}
 
 	if (enabled_qtypes(c)) {
@@ -1203,23 +1267,21 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 	mutex_lock(&c->sb_lock);
 	if (c->opts.version_upgrade) {
-		if (c->sb.version < bcachefs_metadata_version_new_versioning)
-			c->disk_sb.sb->version_min =
-				le16_to_cpu(bcachefs_metadata_version_min);
-		c->disk_sb.sb->version = le16_to_cpu(bcachefs_metadata_version_current);
-		c->disk_sb.sb->features[0] |= BCH_SB_FEATURES_ALL;
+		c->disk_sb.sb->version = cpu_to_le16(bcachefs_metadata_version_current);
+		c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_SB_FEATURES_ALL);
 		write_sb = true;
 	}
 
 	if (!test_bit(BCH_FS_ERROR, &c->flags)) {
-		c->disk_sb.sb->compat[0] |= 1ULL << BCH_COMPAT_FEAT_ALLOC_INFO;
+		c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_alloc_info);
 		write_sb = true;
 	}
 
 	if (c->opts.fsck &&
-	    !test_bit(BCH_FS_ERROR, &c->flags)) {
-		c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_atomic_nlink;
+	    !test_bit(BCH_FS_ERROR, &c->flags) &&
+	    !test_bit(BCH_FS_ERRORS_NOT_FIXED, &c->flags)) {
 		SET_BCH_SB_HAS_ERRORS(c->disk_sb.sb, 0);
+		SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, 0);
 		write_sb = true;
 	}
 
@@ -1227,13 +1289,30 @@ int bch2_fs_recovery(struct bch_fs *c)
 		bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
+	if (!(c->sb.compat & (1ULL << BCH_COMPAT_extents_above_btree_updates_done)) ||
+	    !(c->sb.compat & (1ULL << BCH_COMPAT_bformat_overflow_done)) ||
+	    le16_to_cpu(c->sb.version_min) < bcachefs_metadata_version_btree_ptr_sectors_written) {
+		struct bch_move_stats stats;
+
+		bch_move_stats_init(&stats, "recovery");
+
+		bch_info(c, "scanning for old btree nodes");
+		ret = bch2_fs_read_write(c);
+		if (ret)
+			goto err;
+
+		ret = bch2_scan_old_btree_nodes(c, &stats);
+		if (ret)
+			goto err;
+		bch_info(c, "scanning for old btree nodes done");
+	}
+
 	if (c->journal_seq_blacklist_table &&
 	    c->journal_seq_blacklist_table->nr > 128)
 		queue_work(system_long_wq, &c->journal_seq_blacklist_gc_work);
-out:
+
 	ret = 0;
-err:
-fsck_err:
+out:
 	set_bit(BCH_FS_FSCK_DONE, &c->flags);
 	bch2_flush_fsck_errs(c);
 
@@ -1247,6 +1326,10 @@ fsck_err:
 	else
 		bch_verbose(c, "ret %i", ret);
 	return ret;
+err:
+fsck_err:
+	bch2_fs_emergency_read_only(c);
+	goto out;
 }
 
 int bch2_fs_initialize(struct bch_fs *c)
@@ -1263,28 +1346,23 @@ int bch2_fs_initialize(struct bch_fs *c)
 	bch_notice(c, "initializing new filesystem");
 
 	mutex_lock(&c->sb_lock);
-	for_each_online_member(ca, c, i)
-		bch2_mark_dev_superblock(c, ca, 0);
-	mutex_unlock(&c->sb_lock);
-
-	mutex_lock(&c->sb_lock);
-	c->disk_sb.sb->version = c->disk_sb.sb->version_min =
-		le16_to_cpu(bcachefs_metadata_version_current);
-	c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_atomic_nlink;
-	c->disk_sb.sb->features[0] |= BCH_SB_FEATURES_ALL;
+	c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
+	c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
 
-	bch2_write_super(c);
+	if (c->opts.version_upgrade) {
+		c->disk_sb.sb->version = cpu_to_le16(bcachefs_metadata_version_current);
+		c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_SB_FEATURES_ALL);
+		bch2_write_super(c);
+	}
 	mutex_unlock(&c->sb_lock);
 
-	set_bit(BCH_FS_ALLOC_READ_DONE, &c->flags);
 	set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
+	set_bit(BCH_FS_MAY_GO_RW, &c->flags);
+	set_bit(BCH_FS_FSCK_DONE, &c->flags);
 
 	for (i = 0; i < BTREE_ID_NR; i++)
 		bch2_btree_root_alloc(c, i);
 
-	set_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags);
-	set_bit(JOURNAL_RECLAIM_STARTED, &c->journal.flags);
-
 	err = "unable to allocate journal buckets";
 	for_each_online_member(ca, c, i) {
 		ret = bch2_dev_journal_alloc(ca);
@@ -1310,18 +1388,45 @@ int bch2_fs_initialize(struct bch_fs *c)
 	 * Write out the superblock and journal buckets, now that we can do
 	 * btree updates
 	 */
-	err = "error writing alloc info";
-	ret = bch2_alloc_write(c, 0);
+	bch_verbose(c, "marking superblocks");
+	err = "error marking superblock and journal";
+	for_each_member_device(ca, c, i) {
+		ret = bch2_trans_mark_dev_sb(c, ca);
+		if (ret) {
+			percpu_ref_put(&ca->ref);
+			goto err;
+		}
+
+		ca->new_fs_bucket_idx = 0;
+	}
+
+	bch_verbose(c, "initializing freespace");
+	err = "error initializing freespace";
+	ret = bch2_fs_freespace_init(c);
+	if (ret)
+		goto err;
+
+	err = "error creating root snapshot node";
+	ret = bch2_fs_initialize_subvolumes(c);
+	if (ret)
+		goto err;
+
+	bch_verbose(c, "reading snapshots table");
+	err = "error reading snapshots table";
+	ret = bch2_fs_snapshots_start(c);
 	if (ret)
 		goto err;
+	bch_verbose(c, "reading snapshots done");
 
 	bch2_inode_init(c, &root_inode, 0, 0,
 			S_IFDIR|S_IRWXU|S_IRUGO|S_IXUGO, 0, NULL);
-	root_inode.bi_inum = BCACHEFS_ROOT_INO;
+	root_inode.bi_inum	= BCACHEFS_ROOT_INO;
+	root_inode.bi_subvol	= BCACHEFS_ROOT_SUBVOL;
 	bch2_inode_pack(c, &packed_inode, &root_inode);
+	packed_inode.inode.k.p.snapshot = U32_MAX;
 
 	err = "error creating root directory";
-	ret = bch2_btree_insert(c, BTREE_ID_INODES,
+	ret = bch2_btree_insert(c, BTREE_ID_inodes,
 				&packed_inode.inode.k_i,
 				NULL, NULL, 0);
 	if (ret)
@@ -1331,13 +1436,16 @@ int bch2_fs_initialize(struct bch_fs *c)
 
 	err = "error creating lost+found";
 	ret = bch2_trans_do(c, NULL, NULL, 0,
-		bch2_create_trans(&trans, BCACHEFS_ROOT_INO,
+		bch2_create_trans(&trans,
+				  BCACHEFS_ROOT_SUBVOL_INUM,
 				  &root_inode, &lostfound_inode,
 				  &lostfound,
 				  0, 0, S_IFDIR|0700, 0,
-				  NULL, NULL));
-	if (ret)
+				  NULL, NULL, (subvol_inum) { 0 }, 0));
+	if (ret) {
+		bch_err(c, "error creating lost+found");
 		goto err;
+	}
 
 	if (enabled_qtypes(c)) {
 		ret = bch2_fs_quota_read(c);
@@ -1346,7 +1454,7 @@ int bch2_fs_initialize(struct bch_fs *c)
 	}
 
 	err = "error writing first journal entry";
-	ret = bch2_journal_meta(&c->journal);
+	ret = bch2_journal_flush(&c->journal);
 	if (ret)
 		goto err;