bcachefs: bch_sb_field_clean

Implement a superblock field so we don't have to read the journal after
a clean shutdown (and more importantly, we can verify what we find in
the journal after a clean shutdown)

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

7 files changed, 321 insertions, 111 deletions

diff --git a/fs/bcachefs/bcachefs_format.h b/fs/bcachefs/bcachefs_format.h
index dffb758ad834..8e74de4f8c32 100644
--- a/fs/bcachefs/bcachefs_format.h
+++ b/fs/bcachefs/bcachefs_format.h
@@ -878,7 +878,8 @@ struct bch_sb_field {
 	x(crypt,	2)	\
 	x(replicas,	3)	\
 	x(quota,	4)	\
-	x(disk_groups,	5)
+	x(disk_groups,	5)	\
+	x(clean,	6)
 
 enum bch_sb_field_type {
 #define x(f, nr)	BCH_SB_FIELD_##f = nr,
@@ -1048,6 +1049,37 @@ struct bch_sb_field_disk_groups {
 	struct bch_disk_group	entries[0];
 };
 
+/*
+ * On clean shutdown, store btree roots and current journal sequence number in
+ * the superblock:
+ */
+struct jset_entry {
+	__le16			u64s;
+	__u8			btree_id;
+	__u8			level;
+	__u8			type; /* designates what this jset holds */
+	__u8			pad[3];
+
+	union {
+		struct bkey_i	start[0];
+		__u64		_data[0];
+	};
+};
+
+struct bch_sb_field_clean {
+	struct bch_sb_field	field;
+
+	__le32			flags;
+	__le16			read_clock;
+	__le16			write_clock;
+	__le64			journal_seq;
+
+	union {
+		struct jset_entry start[0];
+		__u64		_data[0];
+	};
+};
+
 /* Superblock: */
 
 /*
@@ -1265,19 +1297,6 @@ static inline __u64 __bset_magic(struct bch_sb *sb)
 #define BCACHE_JSET_VERSION_JKEYS	2
 #define BCACHE_JSET_VERSION		2
 
-struct jset_entry {
-	__le16			u64s;
-	__u8			btree_id;
-	__u8			level;
-	__u8			type; /* designates what this jset holds */
-	__u8			pad[3];
-
-	union {
-		struct bkey_i	start[0];
-		__u64		_data[0];
-	};
-};
-
 #define JSET_KEYS_U64s	(sizeof(struct jset_entry) / sizeof(__u64))
 
 #define BCH_JSET_ENTRY_TYPES()			\
diff --git a/fs/bcachefs/journal_io.c b/fs/bcachefs/journal_io.c
index 4f873ccc492a..8a4e7b2a92ce 100644
--- a/fs/bcachefs/journal_io.c
+++ b/fs/bcachefs/journal_io.c
@@ -13,37 +13,6 @@
 
 #include <trace/events/bcachefs.h>
 
-static struct jset_entry *bch2_journal_find_entry(struct jset *j, unsigned type,
-						 enum btree_id id)
-{
-	struct jset_entry *entry;
-
-	for_each_jset_entry_type(entry, j, type)
-		if (entry->btree_id == id)
-			return entry;
-
-	return NULL;
-}
-
-struct bkey_i *bch2_journal_find_btree_root(struct bch_fs *c, struct jset *j,
-					   enum btree_id id, unsigned *level)
-{
-	struct bkey_i *k;
-	struct jset_entry *entry =
-		bch2_journal_find_entry(j, BCH_JSET_ENTRY_btree_root, id);
-
-	if (!entry)
-		return NULL;
-
-	if (!entry->u64s)
-		return ERR_PTR(-EINVAL);
-
-	k = entry->start;
-	*level = entry->level;
-	*level = entry->level;
-	return k;
-}
-
 struct journal_list {
 	struct closure		cl;
 	struct mutex		lock;
@@ -717,6 +686,37 @@ void bch2_journal_entries_free(struct list_head *list)
 	}
 }
 
+int bch2_journal_set_seq(struct bch_fs *c, u64 last_seq, u64 end_seq)
+{
+	struct journal *j = &c->journal;
+	struct journal_entry_pin_list *p;
+	u64 seq, nr = end_seq - last_seq + 1;
+
+	if (nr > j->pin.size) {
+		free_fifo(&j->pin);
+		init_fifo(&j->pin, roundup_pow_of_two(nr), GFP_KERNEL);
+		if (!j->pin.data) {
+			bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
+			return -ENOMEM;
+		}
+	}
+
+	atomic64_set(&j->seq, end_seq);
+	j->last_seq_ondisk = last_seq;
+
+	j->pin.front	= last_seq;
+	j->pin.back	= end_seq + 1;
+
+	fifo_for_each_entry_ptr(p, &j->pin, seq) {
+		INIT_LIST_HEAD(&p->list);
+		INIT_LIST_HEAD(&p->flushed);
+		atomic_set(&p->count, 0);
+		p->devs.nr = 0;
+	}
+
+	return 0;
+}
+
 int bch2_journal_read(struct bch_fs *c, struct list_head *list)
 {
 	struct journal *j = &c->journal;
@@ -724,10 +724,9 @@ int bch2_journal_read(struct bch_fs *c, struct list_head *list)
 	struct journal_replay *i;
 	struct journal_entry_pin_list *p;
 	struct bch_dev *ca;
-	u64 cur_seq, end_seq, seq;
+	u64 cur_seq, end_seq;
 	unsigned iter;
-	size_t entries = 0;
-	u64 nr, keys = 0;
+	size_t keys = 0, entries = 0;
 	bool degraded = false;
 	int ret = 0;
 
@@ -783,43 +782,13 @@ int bch2_journal_read(struct bch_fs *c, struct list_head *list)
 		}
 	}
 
-	list_for_each_entry(i, list, list) {
-		struct jset_entry *entry;
-		struct bkey_i *k, *_n;
-
-		for_each_jset_key(k, _n, entry, &i->j)
-			keys++;
-	}
-
 	i = list_last_entry(list, struct journal_replay, list);
 
-	nr = le64_to_cpu(i->j.seq) - le64_to_cpu(i->j.last_seq) + 1;
-
-	fsck_err_on(c->sb.clean && (keys || nr > 1), c,
-		    "filesystem marked clean but journal not empty (%llu keys in %llu entries)",
-		    keys, nr);
-
-	if (nr > j->pin.size) {
-		free_fifo(&j->pin);
-		init_fifo(&j->pin, roundup_pow_of_two(nr), GFP_KERNEL);
-		if (!j->pin.data) {
-			bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
-			return -ENOMEM;
-		}
-	}
-
-	atomic64_set(&j->seq, le64_to_cpu(i->j.seq));
-	j->last_seq_ondisk = le64_to_cpu(i->j.last_seq);
-
-	j->pin.front	= le64_to_cpu(i->j.last_seq);
-	j->pin.back	= le64_to_cpu(i->j.seq) + 1;
-
-	fifo_for_each_entry_ptr(p, &j->pin, seq) {
-		INIT_LIST_HEAD(&p->list);
-		INIT_LIST_HEAD(&p->flushed);
-		atomic_set(&p->count, 0);
-		p->devs.nr = 0;
-	}
+	ret = bch2_journal_set_seq(c,
+				   le64_to_cpu(i->j.last_seq),
+				   le64_to_cpu(i->j.seq));
+	if (ret)
+		return ret;
 
 	mutex_lock(&j->blacklist_lock);
 
@@ -842,6 +811,8 @@ int bch2_journal_read(struct bch_fs *c, struct list_head *list)
 				struct journal_replay, list)->j.seq);
 
 	list_for_each_entry(i, list, list) {
+		struct jset_entry *entry;
+		struct bkey_i *k, *_n;
 		bool blacklisted;
 
 		mutex_lock(&j->blacklist_lock);
@@ -863,10 +834,13 @@ int bch2_journal_read(struct bch_fs *c, struct list_head *list)
 			journal_last_seq(j), end_seq);
 
 		cur_seq = le64_to_cpu(i->j.seq) + 1;
+
+		for_each_jset_key(k, _n, entry, &i->j)
+			keys++;
 		entries++;
 	}
 
-	bch_info(c, "journal read done, %llu keys in %zu entries, seq %llu",
+	bch_info(c, "journal read done, %zu keys in %zu entries, seq %llu",
 		 keys, entries, journal_cur_seq(j));
 fsck_err:
 	return ret;
diff --git a/fs/bcachefs/journal_io.h b/fs/bcachefs/journal_io.h
index 4236b7fc37ff..e303df9241de 100644
--- a/fs/bcachefs/journal_io.h
+++ b/fs/bcachefs/journal_io.h
@@ -1,9 +1,6 @@
 #ifndef _BCACHEFS_JOURNAL_IO_H
 #define _BCACHEFS_JOURNAL_IO_H
 
-struct bkey_i *bch2_journal_find_btree_root(struct bch_fs *, struct jset *,
-					    enum btree_id, unsigned *);
-
 /*
  * Only used for holding the journal entries we read in btree_journal_read()
  * during cache_registration
@@ -37,6 +34,7 @@ static inline struct jset_entry *__jset_entry_type_next(struct jset *jset,
 	for_each_jset_entry_type(entry, jset, BCH_JSET_ENTRY_btree_keys)	\
 		vstruct_for_each_safe(entry, k, _n)
 
+int bch2_journal_set_seq(struct bch_fs *c, u64, u64);
 int bch2_journal_read(struct bch_fs *, struct list_head *);
 
 int bch2_journal_entry_sectors(struct journal *);
diff --git a/fs/bcachefs/recovery.c b/fs/bcachefs/recovery.c
index 271eca14753d..9881ab2a1e27 100644
--- a/fs/bcachefs/recovery.c
+++ b/fs/bcachefs/recovery.c
@@ -12,11 +12,118 @@
 #include "recovery.h"
 #include "super-io.h"
 
+struct bkey_i *btree_root_find(struct bch_fs *c,
+			       struct bch_sb_field_clean *clean,
+			       struct jset *j,
+			       enum btree_id id, unsigned *level)
+{
+	struct bkey_i *k;
+	struct jset_entry *entry, *start, *end;
+
+	if (clean) {
+		start = clean->start;
+		end = vstruct_end(&clean->field);
+	} else {
+		start = j->start;
+		end = vstruct_last(j);
+	}
+
+	for (entry = start; entry < end; entry = vstruct_next(entry))
+		if (entry->type == BCH_JSET_ENTRY_btree_root &&
+		    entry->btree_id == id)
+			goto found;
+
+	return NULL;
+found:
+	if (!entry->u64s)
+		return ERR_PTR(-EINVAL);
+
+	k = entry->start;
+	*level = entry->level;
+	return k;
+}
+
+static int verify_superblock_clean(struct bch_fs *c,
+				   struct bch_sb_field_clean *clean,
+				   struct jset *j)
+{
+	unsigned i;
+	int ret = 0;
+
+	if (!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),
+			le64_to_cpu(j->seq)))
+		bch2_fs_mark_clean(c, false);
+
+	mustfix_fsck_err_on(j->read_clock != clean->read_clock, c,
+			"superblock read clock doesn't match journal after clean shutdown");
+	mustfix_fsck_err_on(j->write_clock != clean->write_clock, c,
+			"superblock read clock doesn't match journal after clean shutdown");
+
+	for (i = 0; i < BTREE_ID_NR; i++) {
+		struct bkey_i *k1, *k2;
+		unsigned l1, l2;
+
+		k1 = btree_root_find(c, clean, NULL, i, &l1);
+		k2 = btree_root_find(c, NULL, j, i, &l2);
+
+		if (!k1 && !k2)
+			continue;
+
+		if (!k1 || !k2 ||
+		    k1->k.u64s != k2->k.u64s ||
+		    memcmp(k1, k2, bkey_bytes(k1)) ||
+		    l1 != l2)
+			panic("k1 %px l1 %u k2 %px l2 %u\n", k1, l1, k2, l2);
+
+		mustfix_fsck_err_on(!k1 || !k2 ||
+				    k1->k.u64s != k2->k.u64s ||
+				    memcmp(k1, k2, bkey_bytes(k1)) ||
+				    l1 != l2, c,
+			"superblock btree root doesn't match journal after clean shutdown");
+	}
+fsck_err:
+	return ret;
+}
+
+static bool journal_empty(struct list_head *journal)
+{
+	struct journal_replay *i;
+	struct jset_entry *entry;
+
+	if (list_empty(journal))
+		return true;
+
+	i = list_last_entry(journal, struct journal_replay, list);
+
+	if (i->j.last_seq != i->j.seq)
+		return false;
+
+	list_for_each_entry(i, journal, list) {
+		vstruct_for_each(&i->j, entry) {
+			if (entry->type == BCH_JSET_ENTRY_btree_root)
+				continue;
+
+			if (entry->type == BCH_JSET_ENTRY_btree_keys &&
+			    !entry->u64s)
+				continue;
+			return false;
+		}
+	}
+
+	return true;
+}
+
 int bch2_fs_recovery(struct bch_fs *c)
 {
 	const char *err = "cannot allocate memory";
+	struct bch_sb_field_clean *clean = NULL, *sb_clean = NULL;
 	LIST_HEAD(journal);
-	struct jset *j;
+	struct jset *j = NULL;
 	unsigned i;
 	int ret;
 
@@ -25,22 +132,57 @@ int bch2_fs_recovery(struct bch_fs *c)
 		bch_info(c, "building replicas info");
 		set_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
 	}
+
+	if (c->sb.clean)
+		sb_clean = bch2_sb_get_clean(c->disk_sb.sb);
+	if (sb_clean) {
+		clean = kmemdup(sb_clean, vstruct_bytes(&sb_clean->field),
+				GFP_KERNEL);
+		if (!clean) {
+			ret = -ENOMEM;
+			mutex_unlock(&c->sb_lock);
+			goto err;
+		}
+	}
 	mutex_unlock(&c->sb_lock);
 
-	ret = bch2_journal_read(c, &journal);
+	if (clean)
+		bch_info(c, "recovering from clean shutdown, journal seq %llu",
+			 le64_to_cpu(clean->journal_seq));
+
+	if (!clean || !c->opts.nofsck) {
+		ret = bch2_journal_read(c, &journal);
+		if (ret)
+			goto err;
+
+		j = &list_entry(journal.prev, struct journal_replay, list)->j;
+	} else {
+		ret = bch2_journal_set_seq(c,
+					   le64_to_cpu(clean->journal_seq),
+					   le64_to_cpu(clean->journal_seq));
+		BUG_ON(ret);
+	}
+
+	ret = verify_superblock_clean(c, clean, j);
 	if (ret)
 		goto err;
 
-	j = &list_entry(journal.prev, struct journal_replay, list)->j;
+	fsck_err_on(clean && !journal_empty(&journal), c,
+		    "filesystem marked clean but journal not empty");
 
-	c->bucket_clock[READ].hand = le16_to_cpu(j->read_clock);
-	c->bucket_clock[WRITE].hand = le16_to_cpu(j->write_clock);
+	if (clean) {
+		c->bucket_clock[READ].hand = le16_to_cpu(clean->read_clock);
+		c->bucket_clock[WRITE].hand = le16_to_cpu(clean->write_clock);
+	} else {
+		c->bucket_clock[READ].hand = le16_to_cpu(j->read_clock);
+		c->bucket_clock[WRITE].hand = le16_to_cpu(j->write_clock);
+	}
 
 	for (i = 0; i < BTREE_ID_NR; i++) {
 		unsigned level;
 		struct bkey_i *k;
 
-		k = bch2_journal_find_btree_root(c, j, i, &level);
+		k = btree_root_find(c, clean, j, i, &level);
 		if (!k)
 			continue;
 
@@ -75,15 +217,17 @@ int bch2_fs_recovery(struct bch_fs *c)
 		goto err;
 	bch_verbose(c, "mark and sweep done");
 
-	mutex_lock(&c->sb_lock);
-	SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
-	bch2_write_super(c);
-	mutex_unlock(&c->sb_lock);
-
 	if (c->opts.noreplay)
 		goto out;
 
 	/*
+	 * Mark dirty before journal replay, fsck:
+	 * XXX: after a clean shutdown, this could be done lazily only when fsck
+	 * finds an error
+	 */
+	bch2_fs_mark_clean(c, false);
+
+	/*
 	 * bch2_fs_journal_start() can't happen sooner, or btree_gc_finish()
 	 * will give spurious errors about oldest_gen > bucket_gen -
 	 * this is a hack but oh well.
@@ -121,6 +265,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 out:
 	bch2_journal_entries_free(&journal);
+	kfree(clean);
 	return ret;
 err:
 fsck_err:
diff --git a/fs/bcachefs/super-io.c b/fs/bcachefs/super-io.c
index 9772d5973078..54de9fac6e22 100644
--- a/fs/bcachefs/super-io.c
+++ b/fs/bcachefs/super-io.c
@@ -4,6 +4,7 @@
 #include "disk_groups.h"
 #include "error.h"
 #include "io.h"
+#include "journal.h"
 #include "replicas.h"
 #include "quota.h"
 #include "super-io.h"
@@ -89,6 +90,9 @@ int bch2_sb_realloc(struct bch_sb_handle *sb, unsigned u64s)
 	struct bch_sb *new_sb;
 	struct bio *bio;
 
+	if (sb->sb && sb->page_order >= order)
+		return 0;
+
 	if (sb->have_layout) {
 		u64 max_bytes = 512 << sb->sb->layout.sb_max_size_bits;
 
@@ -849,6 +853,84 @@ static const struct bch_sb_field_ops bch_sb_field_ops_crypt = {
 	.validate	= bch2_sb_validate_crypt,
 };
 
+/* BCH_SB_FIELD_clean: */
+
+void bch2_fs_mark_clean(struct bch_fs *c, bool clean)
+{
+	struct bch_sb_field_clean *sb_clean;
+	unsigned u64s = sizeof(*sb_clean) / sizeof(u64);
+	struct jset_entry *entry;
+	struct btree_root *r;
+
+	mutex_lock(&c->sb_lock);
+	if (clean == BCH_SB_CLEAN(c->disk_sb.sb))
+		goto out;
+
+	SET_BCH_SB_CLEAN(c->disk_sb.sb, clean);
+
+	if (!clean)
+		goto write_super;
+
+	mutex_lock(&c->btree_root_lock);
+
+	for (r = c->btree_roots;
+	     r < c->btree_roots + BTREE_ID_NR;
+	     r++)
+		if (r->alive)
+			u64s += jset_u64s(r->key.u64s);
+
+	sb_clean = bch2_sb_resize_clean(&c->disk_sb, u64s);
+	if (!sb_clean) {
+		bch_err(c, "error resizing superblock while setting filesystem clean");
+		goto out;
+	}
+
+	sb_clean->flags		= 0;
+	sb_clean->read_clock	= cpu_to_le16(c->bucket_clock[READ].hand);
+	sb_clean->write_clock	= cpu_to_le16(c->bucket_clock[WRITE].hand);
+	sb_clean->journal_seq	= journal_cur_seq(&c->journal) - 1;
+
+	entry = sb_clean->start;
+	memset(entry, 0,
+	       vstruct_end(&sb_clean->field) - (void *) entry);
+
+	for (r = c->btree_roots;
+	     r < c->btree_roots + BTREE_ID_NR;
+	     r++)
+		if (r->alive) {
+			entry->u64s	= r->key.u64s;
+			entry->btree_id	= r - c->btree_roots;
+			entry->level	= r->level;
+			entry->type	= BCH_JSET_ENTRY_btree_root;
+			bkey_copy(&entry->start[0], &r->key);
+			entry = vstruct_next(entry);
+			BUG_ON((void *) entry > vstruct_end(&sb_clean->field));
+		}
+
+	BUG_ON(entry != vstruct_end(&sb_clean->field));
+
+	mutex_unlock(&c->btree_root_lock);
+write_super:
+	bch2_write_super(c);
+out:
+	mutex_unlock(&c->sb_lock);
+}
+
+static const char *bch2_sb_validate_clean(struct bch_sb *sb,
+					  struct bch_sb_field *f)
+{
+	struct bch_sb_field_clean *clean = field_to_type(f, clean);
+
+	if (vstruct_bytes(&clean->field) < sizeof(*clean))
+		return "invalid field crypt: wrong size";
+
+	return NULL;
+}
+
+static const struct bch_sb_field_ops bch_sb_field_ops_clean = {
+	.validate	= bch2_sb_validate_clean,
+};
+
 static const struct bch_sb_field_ops *bch2_sb_field_ops[] = {
 #define x(f, nr)					\
 	[BCH_SB_FIELD_##f] = &bch_sb_field_ops_##f,
diff --git a/fs/bcachefs/super-io.h b/fs/bcachefs/super-io.h
index 995b1c907318..7d09d8e45816 100644
--- a/fs/bcachefs/super-io.h
+++ b/fs/bcachefs/super-io.h
@@ -131,6 +131,10 @@ static inline struct bch_member_cpu bch2_mi_to_cpu(struct bch_member *mi)
 	};
 }
 
+/* BCH_SB_FIELD_clean: */
+
+void bch2_fs_mark_clean(struct bch_fs *, bool);
+
 size_t bch2_sb_field_to_text(char *, size_t, struct bch_sb *,
 			     struct bch_sb_field *);
 
diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c
index 6cafbdb888ee..a2a32b924434 100644
--- a/fs/bcachefs/super.c
+++ b/fs/bcachefs/super.c
@@ -199,16 +199,6 @@ int bch2_congested(void *data, int bdi_bits)
  * - allocator depends on the journal (when it rewrites prios and gens)
  */
 
-static void bch2_fs_mark_clean(struct bch_fs *c, bool clean)
-{
-	mutex_lock(&c->sb_lock);
-	if (BCH_SB_CLEAN(c->disk_sb.sb) != clean) {
-		SET_BCH_SB_CLEAN(c->disk_sb.sb, clean);
-		bch2_write_super(c);
-	}
-	mutex_unlock(&c->sb_lock);
-}
-
 static void __bch2_fs_read_only(struct bch_fs *c)
 {
 	struct bch_dev *ca;
@@ -311,10 +301,8 @@ void bch2_fs_read_only(struct bch_fs *c)
 
 	if (!bch2_journal_error(&c->journal) &&
 	    !test_bit(BCH_FS_ERROR, &c->flags) &&
-	    !test_bit(BCH_FS_EMERGENCY_RO, &c->flags)) {
-
+	    !test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
 		bch2_fs_mark_clean(c, true);
-	}
 
 	if (c->state != BCH_FS_STOPPING)
 		c->state = BCH_FS_RO;