libfiemap
is a library for creating block-devices that are backed by storage in read-write partitions. It exists primary for gsid. Generally, the library works by using libfiemap_writer
to allocate large files within filesystem, and then tracks their extents.
There are three main uses for libfiemap
:
system_gsi
image to store Dynamic System Updates.Image creation is done through SplitFiemap
. Depending on the file system, a large image may have to be split into multiple files. On Ext4 the limit is 16GiB and on FAT32 it's 4GiB. Images are saved into /data/gsi/<name>/
where <name>
is chosen by the process requesting the image.
At the same time, a file called /metadata/gsi/<name>/lp_metadata
is created. This is a super partition header that allows first-stage init to create dynamic partitions from the image files. It also tracks the canonical size of the image, since the file size may be larger due to alignment.
It is easy to make block devices out of blocks on /data
when it is not mounted, so first-stage init has no issues mapping dynamic partitions from images. After /data
is mounted however, there are two problems:
/data
is encrypted./dev/block/by-name/data
may be marked as in-use.We break the problem down into three scenarios.
When FDE or metadata encryption is used, /data
is not mounted from /dev/block/by-name/data
. Instead, it is mounted from an intermediate dm-crypt
or dm-default-key
device. This means the underlying device is not marked in use, and we can create new dm-linear devices on top of it.
On these devices, a block device for an image will consist of a single device-mapper device with a dm-linear
table entry for each extent in the backing file.
When a device is unencrypted, or is encrypted with FBE but not metadata encryption, we instead use a loop device with LOOP_SET_DIRECT_IO
enabled. Since /data/gsi
has encryption disabled, this means the raw blocks will be unencrypted as well.
If an image was too large to store a single file on the underlying filesystem, on an FBE/unencrypted device we will have multiple loop devices. In this case, we create a device-mapper device as well. For each loop device it will have one dm-linear
table entry spanning the length of the device.
It's important that we know whether or not an image is currently in-use by a block device. It could be catastrophic to write to a dm-linear device if the underlying blocks are no longer owned by the original file. Thus, when mapping an image, we create a property called gsid.mapped_image.<name>
and set it to the path of the block device.
Additionally, we create a /metadata/gsi/<subdir>/<name>.status
file. Each line in this file denotes a dependency on either a device-mapper node or a loop device. When deleting a block device, this file is used to release all resources.