[PATCH v2] btrfs: block incompatible optional features at scan

[PATCH v2] btrfs: block incompatible optional features at scan

[Anand Jain]

For the matter of completeness we need to check if the device
being scanned has features that are known to the kernel. As of
now if it doesn't - the mount will fails, then what is the point
in having those devices added to the btrfs_fs_devices list at
device_list_add().

So block those devices at scan. Which means the original block at
open_ctee() won't reach in case of device with unsupported feature.
But I am leaving that code as it is, without deleting.

Unit testing:
Create progs with the following changes and mkfs.
--------------------------------------------------
diff --git a/ctree.h b/ctree.h
index 5ab0f4a45a15..a8d86facc045 100644
--- a/ctree.h
+++ b/ctree.h
@@ -480,6 +480,7 @@ struct btrfs_super_block {
 #define BTRFS_FEATURE_INCOMPAT_RAID56          (1ULL << 7)
 #define BTRFS_FEATURE_INCOMPAT_SKINNY_METADATA (1ULL << 8)
 #define BTRFS_FEATURE_INCOMPAT_NO_HOLES                (1ULL << 9)
+#define BTRFS_FEATURE_INCOMPAT_TEST            (1ULL << 10)

 #define BTRFS_FEATURE_COMPAT_SUPP              0ULL

@@ -495,7 +496,8 @@ struct btrfs_super_block {
         BTRFS_FEATURE_INCOMPAT_RAID56 |                \
         BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS |          \
         BTRFS_FEATURE_INCOMPAT_SKINNY_METADATA |       \
-        BTRFS_FEATURE_INCOMPAT_NO_HOLES)
+        BTRFS_FEATURE_INCOMPAT_NO_HOLES |              \
+        BTRFS_FEATURE_INCOMPAT_TEST)

 /*
  * A leaf is full of items. offset and size tell us where to find
diff --git a/mkfs.c b/mkfs.c
index ea584042db16..f3665a93364b 100644
--- a/mkfs.c
+++ b/mkfs.c
@@ -1368,7 +1368,7 @@ int main(int ac, char **av)
        int dev_cnt = 0;
        int saved_optind;
        char fs_uuid[BTRFS_UUID_UNPARSED_SIZE] = { 0 };
-       u64 features = BTRFS_MKFS_DEFAULT_FEATURES;
+       u64 features = BTRFS_MKFS_DEFAULT_FEATURES | BTRFS_FEATURE_INCOMPAT_TEST;
        struct mkfs_allocation allocation = { 0 };
        struct btrfs_mkfs_config mkfs_cfg;
------------------------------------------------------------

Results..

btrfs dev scan /dev/sdc
Scanning for Btrfs filesystems in '/dev/sdc'
ERROR: device scan failed '/dev/sdc' - Protocol family not supported

btrfs dev ready /dev/sdc
ERROR: unable to determine if device '/dev/sdc' is ready for mount: Protocol family not supported

mount /dev/sdc /btrfs
mount: mount /dev/sdc on /btrfs failed: Protocol family not supported

Signed-off-by: Anand Jain <anand.jain@oracle.com>
---

v2: Commit update with unit test case and its results.
    Remove the printk the error to system log, thats not required.
    Return -EPFNOSUPPORT instead of -ENOSUPPORT, thats more appropriate
	when device with incompatible feature is found during device scan.

 fs/btrfs/volumes.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index 6f7873e401ac..8ca3b0d3f1ef 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -1071,6 +1071,7 @@ int btrfs_scan_one_device(const char *path, fmode_t flags, void *holder,
 	u64 transid;
 	u64 total_devices;
 	u64 bytenr;
+	u64 features;
 
 	/*
 	 * we would like to check all the supers, but that would make
@@ -1091,6 +1092,12 @@ int btrfs_scan_one_device(const char *path, fmode_t flags, void *holder,
 	if (btrfs_read_disk_super(bdev, bytenr, &page, &disk_super))
 		goto error_bdev_put;
 
+	features = btrfs_super_incompat_flags(disk_super) &
+				~BTRFS_FEATURE_INCOMPAT_SUPP;
+	if (features) {
+		ret = -EPFNOSUPPORT;
+		goto error_disk_super;
+	}
 	devid = btrfs_stack_device_id(&disk_super->dev_item);
 	transid = btrfs_super_generation(disk_super);
 	total_devices = btrfs_super_num_devices(disk_super);
@@ -1105,6 +1112,7 @@ int btrfs_scan_one_device(const char *path, fmode_t flags, void *holder,
 	if (!ret && fs_devices_ret)
 		(*fs_devices_ret)->total_devices = total_devices;
 
+error_disk_super:
 	btrfs_release_disk_super(page);
 
 error_bdev_put:
-- 
2.7.0

Re: [PATCH v2] btrfs: block incompatible optional features at scan

[David Sterba]

On Wed, Mar 16, 2016 at 08:32:35AM +0800, Anand Jain wrote:
> For the matter of completeness we need to check if the device
> being scanned has features that are known to the kernel. As of
> now if it doesn't - the mount will fails, then what is the point
> in having those devices added to the btrfs_fs_devices list at
> device_list_add().
> 
> So block those devices at scan. Which means the original block at
> open_ctee() won't reach in case of device with unsupported feature.
> But I am leaving that code as it is, without deleting.

I think it makes some sense to skip registration of devices with unknown
features. On the other hand, we'd never be able to test-mount a multiple
device filesystem as the devices won't be in the list. The mount would
fail anyway, but I'd rather keep the decision in one place.

Also, device scan would return a new error condition, so the userspace
tools would need to be updated (not a problem) but older versions with
new kernel will become a bit confusing.

Registration of unsupported device should be silently skipped. Current
state will silently register it, but will never let it past mount, so
it's IMHO OK.