nixpkgs/nixos/lib/make-ext4-fs.nix
Tuomas Tynkkynen bc828721ed nixos/lib/make-ext4-fs: Add a sanity check
I ended up with a corrupted image with the debugfs contraption once, and
given I couldn't reproduce the problem I suppose that happens if the
filesystem of the builder runs out of space.

At least in this instance fsck could detect it, so let's add it as a
sanity check.
2018-05-05 19:30:54 +03:00

94 lines
3.8 KiB
Nix

# Builds an ext4 image containing a populated /nix/store with the closure
# of store paths passed in the storePaths parameter. The generated image
# is sized to only fit its contents, with the expectation that a script
# resizes the filesystem at boot time.
{ pkgs
, storePaths
, volumeLabel
}:
let
sdClosureInfo = pkgs.closureInfo { rootPaths = storePaths; };
in
pkgs.stdenv.mkDerivation {
name = "ext4-fs.img";
nativeBuildInputs = with pkgs; [e2fsprogs.bin libfaketime perl];
buildCommand =
''
# Add the closures of the top-level store objects.
storePaths=$(cat ${sdClosureInfo}/store-paths)
# Also include a manifest of the closures in a format suitable for nix-store --load-db.
cp ${sdClosureInfo}/registration nix-path-registration
# Make a crude approximation of the size of the target image.
# If the script starts failing, increase the fudge factors here.
numInodes=$(find $storePaths | wc -l)
numDataBlocks=$(du -c -B 4096 --apparent-size $storePaths | awk '$2 == "total" { print int($1 * 1.03) }')
bytes=$((2 * 4096 * $numInodes + 4096 * $numDataBlocks))
echo "Creating an EXT4 image of $bytes bytes (numInodes=$numInodes, numDataBlocks=$numDataBlocks)"
truncate -s $bytes $out
faketime -f "1970-01-01 00:00:01" mkfs.ext4 -L ${volumeLabel} -U 44444444-4444-4444-8888-888888888888 $out
# Populate the image contents by piping a bunch of commands to the `debugfs` tool from e2fsprogs.
# For example, to copy /nix/store/abcd...efg-coreutils-8.23/bin/sleep:
# cd /nix/store/abcd...efg-coreutils-8.23/bin
# write /nix/store/abcd...efg-coreutils-8.23/bin/sleep sleep
# sif sleep mode 040555
# sif sleep gid 30000
# In particular, debugfs doesn't handle absolute target paths; you have to 'cd' in the virtual
# filesystem first. Likewise the intermediate directories must already exist (using `find`
# handles that for us). And when setting the file's permissions, the inode type flags (__S_IFDIR,
# __S_IFREG) need to be set as well.
(
echo write nix-path-registration nix-path-registration
echo mkdir nix
echo cd /nix
echo mkdir store
# XXX: This explodes in exciting ways if anything in /nix/store has a space in it.
find $storePaths -printf '%y %f %h %m\n'| while read -r type file dir perms; do
# echo "TYPE=$type DIR=$dir FILE=$file PERMS=$perms" >&2
echo "cd $dir"
case $type in
d)
echo "mkdir $file"
echo sif $file mode $((040000 | 0$perms)) # magic constant is __S_IFDIR
;;
f)
echo "write $dir/$file $file"
echo sif $file mode $((0100000 | 0$perms)) # magic constant is __S_IFREG
;;
l)
echo "symlink $file $(readlink "$dir/$file")"
;;
*)
echo "Unknown entry: $type $dir $file $perms" >&2
exit 1
;;
esac
echo sif $file gid 30000 # chgrp to nixbld
done
) | faketime -f "1970-01-01 00:00:01" debugfs -w $out -f /dev/stdin > errorlog 2>&1
# The debugfs tool doesn't terminate on error nor exit with a non-zero status. Check manually.
if egrep -q 'Could not allocate|File not found' errorlog; then
cat errorlog
echo "--- Failed to create EXT4 image of $bytes bytes (numInodes=$numInodes, numDataBlocks=$numDataBlocks) ---"
return 1
fi
# I have ended up with corrupted images sometimes, I suspect that happens when the build machine's disk gets full during the build.
if ! fsck.ext4 -n -f $out; then
echo "--- Fsck failed for EXT4 image of $bytes bytes (numInodes=$numInodes, numDataBlocks=$numDataBlocks) ---"
cat errorlog
return 1
fi
'';
}