The concept intrigued me, so I actually went on to try your prototype.
I could apply it on v4.12 mainline (newer kernel versions introduce
changes in "struct bio" in "include/linux/blk_types.h" those don't let
the module compile – I think minor changes would be necessary to adapt
to the new struct, though I didn't go into that).
My test scenario:
On a KVM, I created a 64M partition and formatted it to ext4, then put
some random files on it and unmounted the FS. I then called "dmsetup
create bowdev --table "0 131072 bow /dev/vdb1"". The
"/dev/mapper/bowdev" file appeared as expected. I mounted it in
read-only mode ("mount -vo ro /dev/mapper/bowdev /mnt") and run
"fstrim -v /mnt". At this point, I tried to advance to STATE 1 ("echo
1 > /sys/block/dm-2/bow/state"), but I got a kernel BUG alert. The
STATE did not change. I unmounted bowdev and removed the device
("dmsetup remove bowdev") which resulted in 2 subsequent kernel
alerts. The device disappeared but it brought the kernel to an
unstable state (various actions, like sync or trying to recreate the
bow device, resulted in a hang). I could not get any further than
this. I attached all the 3 kernel alerts in "dm-bow.dmesg.log".
This BUG_ON is caused if your file system writes blocks in sizes less
than your page size. I will fix that before I attempt to upstream this
driver assuming it gets accepted. If you can make your file system have
4k blocks, you should be able to proceed (I hit this when I created a
16MB ext4 fs on a loopback device)
I have some questions about dm-bow:
– How file system agnostic this feature is planned to be? While it is
designed with ext4 in mind, is it going to work when used over other
file systems, like FAT or BTRFS for example?
So long as the file system supports fstrim, it should work. If the file
system creates a lot of churn say by running garbage collection, I'd not
recommend it. And I really don't see the use case if the file system has
any sort of snapshot capability - that will always be a superior
solution to a block level one IMO.
– Especially that BTRFS uses a CoW mechanism for even overwriting
files (overwritten segments are written to a free area and only then
gets the old data freed – except some specific conditions when
NO_COW/nodatacow is involved). Won't BTRFS CoW mechanism confuse BoW,
e.g. BTRFS will try to use space that BoW wants to use for backups?
Note however, using BoW on BTRFS wouldn't have much point, since BTRFS
has built-in features for snapshots. This leads me to my next
question.
– Why don't you just use BTRFS on Android? It basically provides a
similar feature like BoW, and it is matured enough, switching
snapshots are easy, etc.. However I see why it wouldn't be feasible
for you, e.g. it is slower than ext4, which would matter for an
Android device.
I'm not the ideal person to answer that question, but yes, I believe
performance is an issue, along with the lack of file based encryption.
– What if you run out of free disk space while updating? I guess you
can just revert to the original state with BoW, but an update might
require more disk space with BoW (and this is a thing, my Android
always complains about not having enough space).
Well this question remains with any snapshot system, and indeed is there
even before you have snapshots. There are really only two choices -
throw away the snapshot and keep going, or fail the update and revert
(with presumably the intent of freeing up more space and trying again.)
Which we choose would be a policy decision - my goal would be to make
sure either option is possible.
– Can I really expect dm-bow to work on non-Android systems (like I
tried it on an Ubuntu KVM)?
Yes, absolutely, but for the moment it's a work in progress and it
contains an assumption about IO accesses being page aligned that is the
reason for the failure you are seeing.
– Do you have any prototype for the command line utility to be used
for recovery?
Yes, and I will be uploading that. For the moment it is embedded in some
Android specific code. It won't take long to extricate it though. It's
actually very simple.
Paul
