The current solution used to collect the list of files installed by
packages does not work for top-level parallel build. Indeed, we rely
on a file created after the installation of the previous package to
build the list of files installed by the current package.
This works well when packages are built sequentially, but badly fails
when using top-level parallel build.
More specifically, top-level parallel build can fail with:
comm: /home/thomas/buildroot/output/build/.files-list-host.new: No such file or directory
Because that file has been removed concurrently by the build process
of another package.
This commit reworks the logic in a very straight-forward way. Before
the installation of each package, we store the list of files that are
already installed and store it in the package build directory. After
the installation of each package, we store again that list of files,
calculate the difference with the before file, and store that as the
list of files installed by that package, still in the package build
directory.
At the end of the build, in target-finalize we collect all the
collected information into the global package file lists, that
continue to be installed in the same location as before, with the same
name.
There are however some differences:
(1) The files are no longer ordered in build order, but by alphabetic
ordering of packages. Indeed, "build order" no longer makes any
sense in the context of top-level parallel build.
(2) Some files which were incorrectly tracked are no longer
tracked. For example, the toolchain package is a target package,
but it installs files in $(HOST_DIR). In the previous logic, the
files installed by the toolchain package in $(HOST_DIR) were
incorrectly affected to the next host package that was installed
after the toolchain package. With our new logic, those files are
no longer tracked at all. To fix this, we would have to change
the logic to scan HOST_DIR/TARGET_DIR/STAGING_DIR for all
installation steps, not just for the install-host, install-target
and install-staging steps respecitively. But the result was
already incorrect anyway, and therefore this should be fixed
separately.
Note that the check_bin_arch hook needs to be adjusted: it was using
the global package-file-list.txt file, but this file is now created
only at the very end of the build. So instead, we use the current
package .file-list.txt file to know which packages have been installed
by the current package in $(TARGET_DIR).
Fixes:
http://autobuild.buildroot.net/results/4e60fa31b1cd08bc7fdf9c5dd3a3f4941e029ba3/
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Commit 509db3b88a added calls to (parts of) the instrumentation steps.
However, those calls are echoed, unlike the other places where we call
them (in the package infra).
Signed-off-by: Yann E. MORIN <yann.morin.1998@free.fr>
Cc: Thomas De Schampheleire <patrickdepinguin@gmail.com>
Acked-by: Thomas De Schampheleire <thomas.de_schampheleire@nokia.com>
Create the staging symlink the same way as the host symlink. This means
using a make dependency rather than recreating it every time.
In coreutils versions below 8.27, re-creation of symbolic links was not
atomic. This means that there is a period in time where the existing link is
removed, before the new one is created. In coreutils 8.27 this was fixed,
see [1]. Note that CentOS 7 ships with coreutils 8.22.
In the following scenario, this is a problem:
- an application is compiled using the sysroot prepared by Buildroot and
links against Xenomai userspace libraries, but its build process is steered
from outside of Buildroot
- to know the correct flags, the application makefile uses the 'xeno-config'
file to request them, and passes DESTDIR=/buildroot/output/staging
- the xeno-config responds with flags based on the path
'/buildroot/output/staging/...'
- while the application build is ongoing, a 'make' happens in Buildroot,
causing the 'staging' symlink to be recreated (even though it already
existed)
- when exactly at this time, the application calls the compiler with -I
flags pointing to output/staging, the build fails with:
-I/buildroot/output/staging/usr/include/xenomai/mercury: Error: ^ is not a directory
-I/buildroot/output/staging/usr/include/xenomai: Error: ^ is not a directory
-I/buildroot/output/staging/usr/include/xenomai/xenomai: Error: ^ is not a directory
-I/buildroot/output/staging/usr/include/xenomai/psos: Error: ^ is not a directory
Failed: ** ^ *
Work around this problem by only creating the staging symlink once, similar
to how the host symlink (if any) is created.
See also commit d0f4f95e39 which changed the
way these symlinks are made. The reasoning in this commit is to move away
from the 'dirs' target.
[1] 376967889e
Signed-off-by: Thomas De Schampheleire <thomas.de_schampheleire@nokia.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
The package instrumentation step 'step_pkg_size' is populating the files:
output/build/packages-file-list.txt
output/build/packages-file-list-staging.txt
output/build/packages-file-list-host.txt
by comparing the list of files before and after installation of a package,
with some clever tricks to detect changes to existing files etc.
As an optimization, instead of gathering this list before and after each
package, where the 'after-state' of one package is the same as the
'before-state' of the next package, only the 'after-state' is used and
is shared between packages.
This works fine, except at the end of the build, as explained next.
In the target-finalize step, many files will be touched. For example, files
like /etc/hosts, /etc/os-release, but also all object files that are
stripped, and all files touched by post-build scripts or created by rootfs
overlays. This means that the 'after-state' of the last package does not
reflect the actual situation after target-finalize is run.
For a single complete build this poses no problem. But, if one incrementally
rebuilds a package after the initial build, e.g. with 'make foo-rebuild',
then all changes that happened in target-finalize at the end of the initial
build (the 'after-state' of the last package built) will be detected as
changes caused by the rebuild of package foo. As a result, all these files
will incorrectly be treated as 'owned' by package foo.
Correct this situation by capturing a new state at the end of
target-finalize, so that the 'before-state' of an incremental build will be
correct.
Note: the reasoning above talks about packages-file-list.txt and
target-finalize, but also applies to
packages-file-list-staging.txt/staging-finalize and
packages-file-list-host.txt/host-finalize.
Signed-off-by: Thomas De Schampheleire <thomas.de_schampheleire@nokia.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
With per-package folder support, top-level parallel build becomes
safe, so we can enclose the .NOTPARALLEL statement in a
!BR2_PER_PACKAGE_DIRECTORIES condition.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Acked-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
This commit implements the core of the move to per-package SDK and
target directories. The main idea is that instead of having a global
output/host and output/target in which all packages install files, we
switch to per-package host and target directories, that only contain
their explicit dependencies.
There are two main benefits:
- Packages will now see only the dependencies they explicitly list in
their <pkg>_DEPENDENCIES variable, and the recursive dependencies
thereof.
- We can support top-level parallel build properly, because a package
only "sees" its own host directory and target directory, isolated
from the build of other packages that can happen in parallel.
It works as follows:
- A new output/per-package/ directory is created, which will contain
one sub-directory per package, and inside it, a "host" directory
and a "target" directory:
output/per-package/busybox/target
output/per-package/busybox/host
output/per-package/host-fakeroot/target
output/per-package/host-fakeroot/host
This output/per-package/ directory is PER_PACKAGE_DIR.
- The global TARGET_DIR and HOST_DIR variable now automatically point
to the per-package directory when PKG is defined. So whenever a
package references $(HOST_DIR) or $(TARGET_DIR) in its build
process, it effectively references the per-package host/target
directories. Note that STAGING_DIR is a sub-dir of HOST_DIR, so it
is handled as well.
- Of course, packages have dependencies, so those dependencies must
be installed in the per-package host and target directories. To do
so, we simply rsync (using hard links to save space and time) the
host and target directories of the direct dependencies of the
package to the current package host and target directories.
We only need to take care of direct dependencies (and not
recursively all dependencies), because we accumulate into those
per-package host and target directories the files installed by the
dependencies. Note that this only works because we make the
assumption that one package does *not* overwrite files installed by
another package.
This is done for "extract dependencies" at the beginning of the
extract step, and for "normal dependencies" at the beginning of the
configure step.
This is basically enough to make per-package SDK and target work. The
only gotcha is that at the end of the build, output/target and
output/host are empty, which means that:
- The filesystem image creation code cannot work.
- We don't have a SDK to build code outside of Buildroot.
In order to fix this, this commit extends the target-finalize step so
that it starts by populating output/target and output/host by
rsync-ing into them the target and host directories of all packages
listed in the $(PACKAGES) variable. It is necessary to do this
sequentially in the target-finalize step and not in each
package. Doing it in package installation means that it can be done in
parallel. In that case, there is a chance that two rsyncs are creating
the same hardlink or directory at the same time, which makes one of
them fail.
This change to per-package directories has an impact on the RPATH
built into the host binaries, as those RPATH now point to various
per-package host directories, and no longer to the global host
directory. We do not try to rewrite such RPATHs during the build as
having such RPATHs is perfectly fine, but we still need to handle two
fallouts from this change:
- The check-host-rpath script, which verifies at the end of each
package installation that it has the appropriate RPATH, is modified
to understand that a RPATH to $(PER_PACKAGE_DIR)/<pkg>/host/lib is
a correct RPAT.
- The fix-rpath script, which mungles the RPATH mainly for the SDK
preparation, is modified to rewrite the RPATH to not point to
per-package directories. Indeed the patchelf --make-rpath-relative
call only works if the RPATH points to the ROOTDIR passed as
argument, and this ROOTDIR is the global host directory. Rewriting
the RPATH to not point to per-package host directories prior to
this is an easy solution to this issue.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Back a few years ago, when we were starting to think about top-level
parallel build, we were not sure how to deal with packages that
installed the same files, so we wanted to catch the situation to assess
how prevalent that was, before we decided what to do and how to address
it.
However, the trend nowadays is that packages will install in a
per-package target/ (and staging/ and host/), and the final directories
will be assembled in a reproducible (alphabetical) order, so if two
packages install the same file, the last one will win (as is currently
the case).
Besides, check-uniq-files reports loads of spurious errors when packages
get reinstalled (e.g. during development).
Finally, check-uniq-files is the only script called during the build,
that is written in python.
So, get rid of check-uniq-files.
Signed-off-by: Yann E. MORIN <yann.morin.1998@free.fr>
Cc: Arnout Vandecappelle <arnout@mind.be>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: Peter Korsgaard <peter@korsgaard.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Currently, we group packages that contribute less then 1%, into the
"Other" category.
However, in some cases, there can be a lot of very comparatively small
packages, and they may not exceed this limit, and so only the "Others"
category would be displayed, which is not nice.
Conversely, if there are a lot of packages, most of which only so
slightly exceeding this limit, then we get all of them in the graph,
which is not nice either.
Add a way for the developers to pass a different cut-off limit. As for
the dependency graph which has BR2_GRAPH_DEPS_OPTS, add the environment
variable BR2_GRAPH_SIZE_OPTS to carry those extra option (in preparation
for more to come, later).
Signed-off-by: Yann E. MORIN <yann.morin.1998@free.fr>
Cc: Thomas De Schampheleire <thomas.de_schampheleire@nokia.com>
[Arnout:
- remove empty base class definition from Config;
- use parser.error instead of ValueError for invalid argument.]
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
When no filesystem is enabled, the $BINARIES_DIR is not created. Yet,
the post-image scripts are still run. When those want to generate an
image in there, they may fail as the dirctory does not exist (it did
exist before we started applying preparatory changes for top-level
parallel build, so scripts got to rely on that assumption).
Do in target-post-image as we do in the sdk rule: create the directory
before calling the scripts.
Signed-off-by: Brent Generous <bgenerous@impinj.com>
[yann.morin.1998@free.fr:
- create the directory before calling the scripts
- don't drop the creation in the sdk rule
]
Signed-off-by: Yann E. MORIN <yann.morin.1998@free.fr>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
This rule was added back in 9429e7b698 (core: introduce an intermediate
rule before the configurators) when the kconfig-side br2-external file
was generated separately from the Makefile-side one.
Now that they are generated together very early in the Makefile, we no
longer need this intermediate rule. Drop it.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Vadim Kochan <vadim4j@gmail.com>
[Peter: also drop outdated reference in the manual]
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
When we introduced support for multiple br2-external trees, we
introduced two files, one on the Makefile side, needed very early,
and one on the kconfig side, needed later in the configuration
process. We naturally introduced a two-step generation, as it looked
like the simplest and most obvious way.
But now, we are on the verge of generating more files on the kconfig
side, and it does not make sense to add even more steps to generate
them.
And even better yet, we can generate both the Makefile-side and
kconfig-side files at the same time, in fact.
Make it so.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Vadim Kochan <vadim4j@gmail.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Now that all the br2-external generated files are named after the same
pattern, it gets easier to remove them all using a glob.
Furthermore, we're on the verge of introducing more such generated
files, so removing them at one fell swoop will be simpler too.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Vadim Kochan <vadim4j@gmail.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Now that the two (all of them!) br2-external related files are generated
in the same location, it makes sense they are named after the same
pattern.
When initial support for (then single) br2-external trees was added back
in a4239f7fd1 (core: introduce the BR2_EXTERNAL variable), it was not
clear-cut why that file was not named with a br2 prefix.
So rename it now.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Vadim Kochan <vadim4j@gmail.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Currently, that file is generated rather late in the configuration
process, so BUILD_DIR is known (and exists) by then.
We're soon to generate that file much earlier, at a point where
BUILD_DIR is not yet known, so we have two options:
1- declare BUILD_DIR earlier;
2- generate the file in an already-known location.
We go with the second solution, as we're already generating a
br2-external related file in BASE_DIR, so we can as well generate all
br2-external files in the same place.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Vadim Kochan <vadim4j@gmail.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
We export GZIP = -n so that GZIP does not record original
name and timestamps. However..
GZIP environment variable is deprecated and soon will not be
supported in future GZIP versions. GZIP suggests the use of a
wrapper to pass options globally but it might be difficult to
implement in Buildroot. For now, we don't export the variable
and fix reproducibility issues per package as they show up in
Autobuilder.
Signed-off-by: Atharva Lele <itsatharva@gmail.com>
Acked-by: Yann E. MORIN <yann.morin.1998@free.fr>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
show-dependency-tree was introduced in this release cycle, as a way to
quickly and easily provide the dependency tree to graph-depends.
show-dependency-tree is no longer used, now that graph-depends has been
switched over to using the more versatile show-info.
Beside, show-dependency-tree has never been part of a release.
Drop it.
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: Thomas De Schampheleire <patrickdepinguin@gmail.com>
Cc: Arnout Vandecappelle <arnout@mind.be>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Sometimes, it is need to quickly get the metadata of a subset of
packages, without resorting to a full-blown JSON query.
Introduce a new per-package (and per-filesystem) foo-show-info rule,
that otputs a per-entity valid JSON blob.
Note that calling it for multiple packages and.or filesystems at once
will not generate a valid JSON blob, as there would be no separator
between the JSON elements:
$ make {foo,bar}-show-info
{ "foo": { foo stuff } }
{ "bar": { bar stuff } }
However, jq is able to absorb this, with its slurping ability, which
generates an array (ellipsed and manualy reformated for readability):
$ make {foo,bar}-show-info |jq -s . -
[
{ "foo": { foo stuff } },
{ "bar": { bar stuff } }
]
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: Thomas De Schampheleire <patrickdepinguin@gmail.com>
Cc: Arnout Vandecappelle <arnout@mind.be>
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Users are increasingly trying to extract information about packages. For
example, they might need to get the list of URIs, or the dependencies of
a package.
Although we do have a bunch of rules to generate some of that, this is
done in ad-hoc way, with most of the output formats just ad-hoc, raw,
unformatted blurbs, mostly internal data dumped as-is.
Introduce a new rule, show-info, that provides a properly formatted
output of all the meta-information about packages: name, type, version,
licenses, dependencies...
We choose to use JSON as the output format, because it is pretty
versatile, has parsers in virtually all languages, has tools to parse
from the shell (jq). It also closely matches Python data structure,
which makes it easy to use with our own internal tools as well. Finally,
JSON being a key-value store, allows for easy expanding the output
without requiring existing consumers to be updated; new, unknown keys
are simply ignored by those (as long as they are true JSON parsers).
The complex part of this change was the conditional output of parts of
the data: virtual packages have no source, version, license or
downloads, unlike non-virtual packages. Same goes for filesystems. We
use a wrapper macro, show-info, that de-multiplexes unto either the
package-related- or filesystem-related macros, and for packages, we also
use a detailed macro for non-virtual packages.
It is non-trivial to properly output correct JSON blurbs, especially
when trying to output an array of objects, like so, where the last item
shall not be followed by a comma: [ { ... }, { ... } ]
So, we use a trick (as sugegsted by Arnout), to $(subst) any pair of
",}" or ", }" or ",]" or ", ]" with only the respective closing symbol,
"}" or "]".
The whole stuff is $(strip)ed to make it a somewhat-minified JSON blurb
that fits on a single line with all spaces squashed (but still with
spaces, as it is not possible to differentiate spaces between JSON
elements from spaces inside JSON strings).
Reported-by: Thomas De Schampheleire <patrickdepinguin@gmail.com>
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Thomas De Schampheleire <patrickdepinguin@gmail.com>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: Arnout Vandecappelle <arnout@mind.be>
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Commit 15cb98769e (release: remove manual build files from release
tarballs) tried to remove the temporary files from the manual build from the
release tarball, but manual-clean only removes build/docs/manual and leaves
build/docs in the tarball.
Instead use 'make clean' to completely remove the build directory from the
tarball.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Currently, when we need to build the full dependency graph, we call make
to show the list of packages (make show-targets), and then call it again
and again iteratively while it returns new packages.
Since calling make will parse the whole set of our Makefiles, this takes
quite a bit of time (~4s each here), and the total can get pretty long.
However, make being make, already builds the whole dependency tree
information, so we can just ask for it.
Add a new top-level rule 'show-dependency-tree' that displays the whole
set of dependencies for all packages. For each package, its name, type
and version is displayed, then all the direct, first-level dependencies
are dumped. We choose a format that is not unlike the dot-graph format,
because it is both easy to read as a human, and easy to parse as a
machine:
foo: target 1.2.3
foo -> bar host-meh
bar: target virtual
bar -> buz
buz: target 2.3.4
buz ->
host-meh: host virtual
host-meh -> host-bleark
host-bleark: host 3.4.5
host-bleark ->
rootfs-meh: host
rootfs-meh -> host-bleark
To be noted: rootfs are currently reported as if they were 'host'
packages, to stay aligned with how graph-depends currently treats them.
Ideally, graph-depends could be enhanced to recognise them separately,
but that is another story.
For just plain defconfig, which is about the smallest config we can have
with an internal toolchain, we already have a seven-fold improvement
(with the graph-depends rule modified to not run the pdf generation, to
be able to just compare the tree generation):
$ time make graph-depends
real 0m27.344s
$ time make show-dependency-tree
real 0m3.848s
>From defconfig, C++, wchar, locales, ssp, and allyespackageconfig,
tweaked for even more packages (qt5 not qt4, luajit to avoid multi
providers, etc...), the timings are (graph-depends still modified to
not generate the pdf):
$ time make graph-depends
real 1m56.459s
$ time make show-dependency-tree
real 0m5.748s
There. I don't think those numbers need any explanation whatsoever;
they do speak on their own. OK, for maths sake, the ratio is about
twenty-fold. So, "yeah", I guess... ;-)
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: Thomas De Schampheleire <patrickdepinguin@gmail.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Using 'make printvars' for printing all variables is not very useful.
E.g. all macros will output some bogus value. In addition, the same can
be achieved with 'make -p'.
We can simply remove the condition on $(VARS). If VARS is not set, the
filter expression will be empty which matches nothing, so nothing is
printed.
Note that the old behaviour can still be achieved with:
make printvars VARS=%
Update the 'make help' text to match the new behaviour.
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Cc: Yann E. MORIN <yann.morin.1998@free.fr>
Tested-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Acked-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Since commit 0db34529f4 we use rsync with the --keep-dirlinks option to
prevent overlays from accidentally overwriding /{usr,bin,sbin,lib} links
when BR2_ROOTFS_MERGED_USR option is enabled. Unfortunately this also
prevents replacing a symlink by a directory on purpose (e.g. /var/log,
to persist system logs).
Steps to reproduce:
- enable BR2_ROOTFS_MERGED_USR and BR2_PACKAGE_SKELETON_INIT_SYSV
- mkdir some_path/rootfs-overlay/var/log
- enable BR2_ROOTFS_OVERLAY="some_path/rootfs-overlay"
- run 'make'
- 'target/var/log' is still a symlink to '../tmp', not a directory
The --keep-dirlinks option can be dropped, since we run sanity checks
on overlays. Now the rsync invocation is identical to the SYSTEM_RSYNC
logic we have in system/system.mk, so use that variable.
Signed-off-by: Carlos Santos <casantos@datacom.ind.br>
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
