Turned out that setting a nil-UUID is no better than clearing it.
What currently happens is as follows:
- first, genext2fs does not generate a UUID
- then we tune2fs to upgrade the filesystem
- then we run fsck, which generates a random UUID
- then we re-run tune2fs to set a nil-UUID
So, on the target, if the file system is improperly unmounted (eg.
with a power failure), on next boot, fsck may be run, and a new
random UUID will be generated.
*However*, fsck improperly updates the filesystem when it adds the
UUID, and there are a few group descriptor checksum errors.
Those errors will go undetected until the next fsck, which will then
block for user input (bad on embedded systems, bad).
Fix that by systematically generating a random UUID _before_ we call
to fsck.
A random UUID is not so bad, after all, since there are already so
many sources of unpredictability in the filesystem: files date and
ordering, files content (date, paths...) which renders a fixed UUID
unneeded.
And it is still possible to set the UUID in a post-image script if
needed, anyway.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Peter Korsgaard <jacmet@uclibc.org>
Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
"tune2fs -U clear" creates an invalid filesystem, that fsck.ext2
whines about later:
$ make rootfs-ext2
[--SNIP--]
$ ./host/usr/sbin/fsck.ext4 images/rootfs.ext2
e2fsck 1.42.7 (21-Jan-2013)
Filesystem did not have a UUID; generating one.
images/rootfs.ext2: clean, 4616/5120 files, 53171/131072 blocks
$ ./host/usr/sbin/fsck.ext4 -f images/rootfs.ext2
e2fsck 1.42.7 (21-Jan-2013)
One or more block group descriptor checksums are invalid. Fix<y>? yes
Group descriptor 0 checksum is 0x4131, should be 0x8bdb. FIXED.
[--SNIP--]
So we set an explicitly NULL UUID instead.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
A quick test with a ~3.5MB ext4 filesystem shows that 1081 blocks isn't
enough:
tune2fs 1.42.7 (21-Jan-2013)
Please run e2fsck on the filesystem.
Creating journal inode:
Journal size too big for filesystem.
So bump it a bit. Overestimating the journal size is probably not really
a big deal for the kind of systems using ext3/4 anyway.
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
Set the count- and time-based checks intervals to 0, thus effectively
disabling automatic checks at boot (after a suggestion by Arnout).
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Arnout Vandecappelle <arnout@mind.be>
Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
Some bootloaders have a buggy ext2 support, and require ext2 rev1
instead of the traditional ext2 rev0 that genext2fs produces.
tune2fs accepts only one '-O list' at a time, so we need to construct
a list of -O options.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Arnout Vandecappelle <arnout@mind.be>
Acked-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
Use the host-e2fsprogs to tune2fs the generated rootfs.ext2 image,
and upgrade it to either one of ext2, ext3 or ext4.
Since calling tune2fs may require running e2fsck (tune2fs will warn
to do so when certain FS options are changed), we systematically call
e2fsck. This makes the code path simpler, and as a side-effect checks
that genext2fs did not generate garbage.
In turn, e2fsck will unconditionally add a UUID to the filesystem,
which is bad for reproducibility, so we call tune2fs again to remove
the UUID. This does not require checking the filesystem.
To ensure compatibility of Buildroot's .config, leave ext2 as the
default. Boards' .config can override this at will.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Arnout Vandecappelle <arnout@mind.be>
Cc: Peter Korsgaard <jacmet@uclibc.org>
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
Closes#2929
Instead of just adding a fixed amount to the blocks used, try to
estimate the real space needed according to the filesystem structure
(bitmaps, inodes, blocks).
The side effect of this is that we no longer significantly overestimate
the size needed for small file systems.
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
With the ROOTFS_TARGET conversion, EXT2_OPTS gets evaluated very early
(before TARGET_DIR is populated with files), so the calculated
blocks/inodes numbers are wrong.
Fix it by moving the calculation to a shell script wrapper around
genext2fs, so it only gets executed just before genext2fs runs.
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
