1 files changed, 76 insertions, 40 deletions

diff --git a/c_src/cmd_migrate.c b/c_src/cmd_migrate.c
index aa17a160..4e4fbcba 100644
--- a/c_src/cmd_migrate.c
+++ b/c_src/cmd_migrate.c
@@ -86,6 +86,15 @@ found:
 	return ret;
 }
 
+static void mark_nouse_range(struct bch_dev *ca, u64 sector_from, u64 sector_to)
+{
+	u64 b = sector_to_bucket(ca, sector_from);
+	do {
+		set_bit(b, ca->buckets_nouse);
+		b++;
+	} while (bucket_to_sector(ca, b) < sector_to);
+}
+
 static void mark_unreserved_space(struct bch_fs *c, ranges extents)
 {
 	struct bch_dev *ca = c->devs[0];
@@ -93,17 +102,16 @@ static void mark_unreserved_space(struct bch_fs *c, ranges extents)
 	struct range i;
 
 	for_each_hole(iter, extents, bucket_to_sector(ca, ca->mi.nbuckets) << 9, i) {
-		u64 b;
-
 		if (i.start == i.end)
 			return;
 
-		b = sector_to_bucket(ca, i.start >> 9);
-		do {
-			set_bit(b, ca->buckets_nouse);
-			b++;
-		} while (bucket_to_sector(ca, b) << 9 < i.end);
+		mark_nouse_range(ca, i.start >> 9,
+			round_up(i.end, 1 << 9) >> 9);
 	}
+
+	/* Also be sure to mark the space for the default sb layout */
+	unsigned sb_size = 1U << ca->disk_sb.sb->layout.sb_max_size_bits;
+	mark_nouse_range(ca, 0, BCH_SB_SECTOR + sb_size * 2);
 }
 
 static ranges reserve_new_fs_space(const char *file_path, unsigned block_size,
@@ -279,7 +287,7 @@ static int migrate_fs(const char		*fs_path,
 		.dev		= stat.st_dev,
 		.extents	= extents,
 		.type		= BCH_MIGRATE_migrate,
-		.reserve_start	= roundup((format_opts.superblock_size * 2 + 8) << 9,
+		.reserve_start	= roundup((format_opts.superblock_size * 2 + BCH_SB_SECTOR) << 9,
 					  bucket_bytes(c->devs[0])),
 	};
 
@@ -381,6 +389,28 @@ static void migrate_superblock_usage(void)
 	     "Report bugs to <linux-bcachefs@vger.kernel.org>");
 }
 
+static void add_default_sb_layout(struct bch_sb* sb, unsigned *out_sb_size)
+{
+	unsigned sb_size = 1U << sb->layout.sb_max_size_bits;
+	if (out_sb_size)
+		*out_sb_size = sb_size;
+
+	if (sb->layout.nr_superblocks >= ARRAY_SIZE(sb->layout.sb_offset))
+		die("Can't add superblock: no space left in superblock layout");
+
+	for (unsigned i = 0; i < sb->layout.nr_superblocks; i++)
+		if (le64_to_cpu(sb->layout.sb_offset[i]) == BCH_SB_SECTOR ||
+		    le64_to_cpu(sb->layout.sb_offset[i]) == BCH_SB_SECTOR + sb_size)
+			die("Superblock layout already has default superblocks");
+
+	memmove(&sb->layout.sb_offset[2],
+		&sb->layout.sb_offset[0],
+		sb->layout.nr_superblocks * sizeof(u64));
+	sb->layout.nr_superblocks += 2;
+	sb->layout.sb_offset[0] = cpu_to_le64(BCH_SB_SECTOR);
+	sb->layout.sb_offset[1] = cpu_to_le64(BCH_SB_SECTOR + sb_size);
+}
+
 int cmd_migrate_superblock(int argc, char *argv[])
 {
 	static const struct option longopts[] = {
@@ -414,34 +444,23 @@ int cmd_migrate_superblock(int argc, char *argv[])
 	if (!sb_offset)
 		die("Please specify offset of existing superblock");
 
-	int fd = xopen(devs.data[0], O_RDWR);
+	int fd = xopen(devs.data[0], O_RDWR | O_EXCL);
 	struct bch_sb *sb = __bch2_super_read(fd, sb_offset);
-	unsigned sb_size = 1U << sb->layout.sb_max_size_bits;
+	unsigned sb_size;
+	/* Check for invocation errors early */
+	add_default_sb_layout(sb, &sb_size);
 
-	if (sb->layout.nr_superblocks >= ARRAY_SIZE(sb->layout.sb_offset))
-		die("Can't add superblock: no space left in superblock layout");
-
-	for (unsigned i = 0; i < sb->layout.nr_superblocks; i++)
-		if (le64_to_cpu(sb->layout.sb_offset[i]) == BCH_SB_SECTOR ||
-		    le64_to_cpu(sb->layout.sb_offset[i]) == BCH_SB_SECTOR + sb_size)
-			die("Superblock layout already has default superblocks");
-
-	memmove(&sb->layout.sb_offset[2],
-		&sb->layout.sb_offset[0],
-		sb->layout.nr_superblocks * sizeof(u64));
-	sb->layout.nr_superblocks += 2;
-	sb->layout.sb_offset[0] = cpu_to_le64(BCH_SB_SECTOR);
-	sb->layout.sb_offset[1] = cpu_to_le64(BCH_SB_SECTOR + sb_size);
+	/* Rewrite first 0-3.5k bytes with zeroes, ensuring we blow away
+	 * the old superblock */
+	// TODO: fix the "Superblock write was silently dropped" warning properly
+	static const char zeroes[(BCH_SB_SECTOR << 9) + sizeof(struct bch_sb)];
+	xpwrite(fd, zeroes, ARRAY_SIZE(zeroes), 0, "zeroing start of disk");
 
-	/* also write first 0-3.5k bytes with zeroes, ensure we blow away old
-	 * superblock */
-	static const char zeroes[BCH_SB_SECTOR << 9];
-	xpwrite(fd, zeroes, BCH_SB_SECTOR << 9, 0, "zeroing start of disk");
-
-	bch2_super_write(fd, sb);
 	xclose(fd);
 
-	/* mark new superblocks */
+	/* We start a normal FS instance with the sb buckets temporarily
+	 * prohibited from allocation, performing any recovery/upgrade/downgrade
+	 * as needed, and only then change the superblock layout */
 
 	struct bch_opts opts = bch2_opts_empty();
 	opt_set(opts, nostart,	true);
@@ -454,29 +473,46 @@ int cmd_migrate_superblock(int argc, char *argv[])
 		die("error opening filesystem: %s", bch2_err_str(ret));
 
 	struct bch_dev *ca = c->devs[0];
-	for (u64 b = 0; bucket_to_sector(ca, b) < BCH_SB_SECTOR + sb_size * 2; b++)
-		set_bit(b, ca->buckets_nouse);
+	mark_nouse_range(ca, 0, BCH_SB_SECTOR + sb_size * 2);
 
 	ret = bch2_fs_start(c);
 	if (ret)
 		die("Error starting filesystem: %s", bch2_err_str(ret));
 
+	BUG_ON(1U << ca->disk_sb.sb->layout.sb_max_size_bits != sb_size);
+
+	/* Here the FS is already RW.
+	 * Apply the superblock layout changes first, everything else can be
+	 * repaired on a subsequent recovery */
+	add_default_sb_layout(ca->disk_sb.sb, NULL);
+	ret = bch2_write_super(c);
+	if (ret)
+		die("Error writing superblock: %s", bch2_err_str(ret));
+
+	/* Now explicitly mark the new sb buckets in FS metadata */
+	ret = bch2_trans_mark_dev_sb(c, ca, BTREE_TRIGGER_transactional);
+	if (ret)
+		die("Error marking superblock buckets: %s", bch2_err_str(ret));
+
 	bch2_fs_stop(c);
 
+#if CONFIG_BCACHEFS_DEBUG
+	/* Verify that filesystem is clean and consistent */
+
 	opts = bch2_opts_empty();
 	opt_set(opts, fsck, true);
 	opt_set(opts, fix_errors, true);
-
-	/*
-	 * Hack: the free space counters are coming out wrong after marking the
-	 * new superblock, but it's just the device counters so it's
-	 * inconsequential:
-	 */
+	opt_set(opts, nochanges, true);
 
 	c = bch2_fs_open(&devs, &opts);
 	ret =   PTR_ERR_OR_ZERO(c);
 	if (ret)
-		die("error opening filesystem: %s", bch2_err_str(ret));
+		die("error checking filesystem: %s", bch2_err_str(ret));
+
+	if (test_bit(BCH_FS_errors, &c->flags) || test_bit(BCH_FS_errors_fixed, &c->flags))
+		die("Filesystem has errors after migration");
+
 	bch2_fs_stop(c);
+#endif
 	return 0;
 }