After a few weeks of running reproducible builds in the autobuilders,
we found out that such builds without a diffoscope analysis are pretty
useless: the cmp on the tarballs doesn't help us fix the
reproducibility issue.
So, let's only do reproducible builds when diffoscope is available.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Tested-by: Atharva Lele <itsatharva@gmail.com>
Acked-by: Atharva Lele <itsatharva@gmail.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
In a future patch, we are going to need a SystemInfo() instance in
gen_config(), so create the SystemInfo() instance there, and pass it
to fixup_config(), where we currently use it.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Tested-by: Atharva Lele <itsatharva@gmail.com>
Acked-by: Atharva Lele <itsatharva@gmail.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Initial patch to start reproducibility testing.
Also enable tar filesystem to compare images since diffoscope works well
with tar.
Signed-off-by: Atharva Lele <itsatharva@gmail.com>
Reviewed-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Now that Qt4 is removed, these fixups can be removed as well.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Currently all random defconfigs which are used in autobuilder use size
optimizaion (-Os), since BR2_OPTIMIZE_S=y is the default.
Adding "-O2" optimization will give better test coverage.
In many cases software gets built with speed optimization rather than
size optimization. So let's add Level 2 optimizaion option to be
generated in random defconfigs, so we could be able to test how
packages are built with "-O2" in autobuilder.
Signed-off-by: Evgeniy Didin <Evgeniy.Didin@synopsys.com>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: arc-buildroot@synopsys.com
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
When adding the custom BR2_WGET value in the configuration,
genrandconfig forgets to add a newline. Due to this, the next option
that is added is printed on the same line as BR2_WGET="", which causes
it to be ignored.
Due to this, in all builds, the line right after BR2_WGET was
ignored. It could have been BR2_ENABLE_DEBUG, BR2_INIT_BUSYBOX,
BR2_INIT_SYSTEMD, BR2_ROOTFS_DEVICE_CREATION_DYNAMIC_EUDEV,
BR2_STATIC_LIBS or BR2_PACKAGE_PYTHON_PY_ONLY depending on the
randomization.
Fix that by adding a proper newline at the end of the BR2_WGET option.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
A number of autobuilder failures are due to the fact that autobuilder
instances use old distributions, with old SSL certificates, and
therefore wget aborts with an error "The certificate of `xyz.org' is
not trusted.".
In order to avoid such failures that are not very interesting in the
context of the autobuilders, we pass --no-check-certificate to
wget. The integrity of the downloaded files is anyway verified by the
hashes, and this is only meant to be used in the context of
testing/CI, not in production.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
In commit ea9669fffa ("core: kill
DEPENDENCIES_HOST_PREREQ"), the core-dependencies make target was
removed, and is now named just "dependencies".
This broke the utils/genrandconfig script, and this commit intends to
fix that.
Since this script is part of the Buildroot tree, it is provided
together with Buildroot, so we don't need to support the legacy
core-dependencies target. Someone checking out an older Buildroot will
have a working setup, with support/dependencies exposing
core-dependencies and utils/genrandconfig using core-dependencies. The
only broken situation will be between
ea9669fffa and this commit, but that's
not a sufficient reason to add some backward compatibility code.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Fix these warnings:
E201 whitespace after '['
E202 whitespace before ']'
E501 line too long (138 > 132 characters)
-> isolate the common part of the external toolchain url in a variable
to make the long lines shorter and more readable.
Signed-off-by: Ricardo Martincoski <ricardo.martincoski@gmail.com>
Cc: Arnout Vandecappelle <arnout@mind.be>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
When generating a random configuration, also check the host
dependencies for that configuration and error out if it fails.
This is mainly intended for the autobuilders, so they don't try to
build a configuration for which the necessary host packages are not
installed.
Instead of erroring out, we could also generate a different random
config. However:
- When called for the autobuilders, the autobuild-run script will
anyway just attempt another build. The only difference is that a
warning will be printed in this autobuild log now. This warning
is in fact quite useful to hint that the autobuilder could benefit
from intalling this additional host package.
- When called manually, the user may in fact want to install the
missing host package and continue with the saved random config.
Fixes:
http://autobuild.buildroot.net/results/d7c/d7c12e873a71fef1e0b5e19cfad45024050614e1http://autobuild.buildroot.net/results/a0a/a0a105c1ec2d0517112b02932a85831d4b597b9fhttp://autobuild.buildroot.net/results/96c/96cc6cedc64c777e993671204809599a1498472a
and many others...
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
In preparation for the addition of comments in the CSV file listing
toolchain configurations, we filter out such lines when reading the
CSV file in utils/genrandconfig.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Reviewed-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Now we have the toolchain config fragments in the buildroot directory
itself, it is no longer necessary to fetch it from the toolchain URL.
The --toolchains-url option is renamed to --toolchains-csv.
The paths in the toolchains_csv file should be either absolute, or
relative to buildrootdir.
After this change, the script should be called from autobuild-run as:
subprocess.call([os.path.join(srcdir, "utils/genrandconfig"),
"-o", outputdir, "-b", srcdir,
"--toolchains-csv", kwargs['toolchains_csv']],
stdout=devnull, stderr=log)
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
This has a number of side-effects which must be handled.
The lines in minimal.config may be overridden by the random lines added
by amending the configuration, so is_toolchain_usable() shouldn't take
those into account, or indeed the random lines added. Therefore, make
a copy of the config before appending minimal.config and the random
lines. While we're at it, rename the variable to the more appropriate
toolchainconfig.
minimal.config sets BR2_INIT_NONE=y, but we really also want to test
with BR2_INIT_BUSYBOX=y. Therefore, add a random line to use the
busybox init system. We set its probability rather high. The
probabilities of systemd and eudev are increased since they're now
in the else branch of BR2_INIT_BUSYBOX, which halves the probability
that we even get there.
We now also generate configurations without busybox. Previously,
busybox was almost always selected due to BR2_INIT_BUSYBOX=y. Only if
systemd is selected there was a chance to build without busybox.
We now set BR2_SYSTEM_BIN_SH_NONE=y, the other /bin/sh options are
never tested. However, this is not really something that is relevant
to test in the autobuilders.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
When outputdir is 'output' (the default), genrandconfig didn't work
correctly because it expects the configfile in outputdir, while
Buildroot puts it in the buildroot directory.
Fix this by explicitly checking if outputdir == buildrootdir/output.
Because abspath is used for both paths, string comparison works
reliably.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
The path to the .config file is calculated in several places - replace
it with a single calculation, and pass configfile as an argument
to is_toolchain_usable and fixup_config. These functions also don't
need outputdir any more.
This makes it easier to fix the case when configfile is not in
outputdir.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
The --instance argument is just an artifact of genrandconfig's
history as part of autobuild-run. It is much more logical to pass
the output directory and the buildroot directory as arguments, with
sane defaults.
This also allows us to remove the hack of creating a symlink in the
instance directory if it doesn't exist yet.
Note that the default outputdir 'output' doesn't work yet, because in
that case Buildroot will put the config file in the buildroot directory
instead of the output directory. This will be fixed in a follow-up
patch.
After this change, the script should be called from autobuild-run as:
subprocess.call([os.path.join(srcdir, "utils/genrandconfig"),
"-o", outputdir, "-b", srcdir,
"--toolchains-url", kwargs['toolchains_url']],
stdout=devnull, stderr=log)
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
This prepares for passing buildrootdir as an argument.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
This prepares for passing outputdir as an argument.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
The output of genrandconfig is currently very terse, which is annoying
for debugging the script or generally seeing what is going on. Also the
timing information added by log_write isn't very useful when the script
is used stand-alone.
In the new setup, (verbose) output goes to stdout and error output goes
to stderr. Also the "INFO: generate the configuration" message is
eliminated - it should go in the autobuild-run script.
We also add an explicit message when a toolchain can't be used after
the first defconfig, otherwise autobuild-run will just silently
restart.
Note that, since the output of make is no longer redirected to
/dev/null, we get one more message on stderr that will be recorded in
the autobuilder's log file: KCONFIG_SEED=0xXXXXXXXX.
This approach allows us to optimise the error handling to use
exceptions, where appropriate, which can be caught at the top level and
converted to an error message to stderr. This, in turn, allows us to use
subprocess.check_call, which eliminates a lot of conditions.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
The SystemInfo class is instantiated globally and passed down to all
functions, but it is really only used in fixup_config. So instead,
instantiate it there.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
kwargs is a left-over from the use of docopt, it's better to use
argparse's Namespace object directly.
In addition, most functions use just one or two fields of args, so
these can just as well be passed directly as arguments to the function.
Particularly for outputdir it doesn't make sense to reconstruct it all
the time.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Warnings fixed:
E731 do not assign a lambda expression, use a def
-> urlopen_closing is defined with a def. urlopen is not used
elsewhere so inlined.
E302 expected 2 blank lines
E501 line too long
-> long lines due to a long string are NOT split
E701 multiple statements on one line (colon)
E722 do not use bare except'
-> use "except Exception", so KeyInterrupt and SystemExit are still
passed. We never intended to catch those.
E741 ambiguous variable name 'l'
-> variable name is replaced with the much more descriptive
toolchains_csv
E271 multiple spaces after keyword
E231 missing whitespace after ','
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Popen is more complicated and more difficult to understand.
check_output raises an exception if the exit code is non-zero, but
that's probably what we want if ldd can't be executed.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
This script will be used by the autobuild-run script to generate the
configuration to test. It is put in the utils directory because it can
also be called directly to allow users to test things.
For now, it is a direct copy of the relevant functions from the
autobuild-run script. The only changes are:
- unneeded import statements are removed;
- code/decode wrappers are limited to decode_byte_list;
- __main__ handling is added.
For now, the only supported arguments are the ones needed for
autobuild-run. Follow-up patches will refactor things and also change
the way the script is called. In this version, it can be called from the
autobuild-run script as:
subprocess.call([os.path.join(srcdir, "utils/genrandconfig"),
"-i", str(kwargs['instance']),
"--toolchains-url", kwargs['toolchains_url']],
stdout=log, stderr=log)
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
