The previous approach used `df` to get usable space and then added a
fixed size to that number in order to account for filesystem
overhead. However, at some point that stopped working for me. It
appears that ext4 filesystem overhead can vary over time or because of
other factors. (Certainly now that I think about it the old code would
only work well for people with the exact same filesystem size as me.)
So the new approach is to just completely ignore what `df` tells us
and instead go directly to the source: the filesystem's internal
notion of exactly how much space it takes up. We use `dumpe2fs` to
retrieve this information and calculate the on-disk size dynamically
from that. Then we add the space that boot data takes up (unchanged),
and we add 5MB padding because when I tested this it didn't quite add
up otherwise. https://unix.stackexchange.com/a/13551/29146 suggests
that this unaccounted-for data may be e.g. additional copies of the
superblock.
See
https://github.com/QubesOS/qubes-core-agent-linux/pull/146#discussion_r238080117
for details; but tl;dr:
* Journal size is 64M
* Inode table is 256 bytes * 643376 inodes allocated = ~157M
* Reserved GDT blocks take up 1024 blocks * 4096 byte block size = 4M
* Fixed-size parts of the filesystem probably take up another MB or
two
These actually adds up to more than the 222M number used in this
commit. But it seems _about_ right, so just Ship It(tm).
This reverts commit 8000e76d43, because
as @marmarek pointed out, the original was correct and I totally
misread. The check in question is checking whether to _abort_, not
whether to continue. So we want to check if the block device size is
_less_ than the filesystem + margin, not more.
ReopensQubesOS/qubes-issues#4553
If xvda is not partitioned, checking size of device mounted as root dev
doesn't detect xvda resize - it will have the new size, while the fs on
it not. Change to checking actual filesystem size.
Also, improve log message to include the size.
FixesQubesOS/qubes-issues#4274
