When running iperf3 with Linux 6.18.51 on VisionFive 2, one CPU core
stays at 100% load during the test, while network throughput is limited
to approximately 793 Mbits/sec:
[ ID] Interval Transfer Bitrate
[ 5] 0.00-10.01 sec 945 MBytes 793 Mbits/sec receiver
The regression was introduced in Linux 6.14 by commit df542f669307
("net: stmmac: Switch to zero-copy in non-XDP RX path"). It leads to
misaligned access exceptions when parsing packet headers. During the
iperf3 test, perf stat reported approximately 9.6 million alignment
faults over 5 seconds:
$ perf stat -a -e alignment-faults -- sleep 5
Performance counter stats for 'system wide':
9620073 alignment-faults
5.029337000 seconds time elapsed
The kernel must emulate these unaligned accesses, resulting in
significant overhead.
The issue was fixed in Linux 6.18.52 [1]. Bump Linux to 6.18.55, the
latest stable release at the time of writing, to include the fix.
[1] https://lore.kernel.org/all/2026091704-CVE-2026-90059-36e6@gregkh
Fixes: 98aa14a186 ("configs/visionfive2: bump Linux to 6.18.51, U-Boot to 2026.07, and OpenSBI to 1.8.1")
Assisted-by: LLM
Signed-off-by: Fengwei Tan <tfx2001@outlook.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Codeberg has changed the way it generate hashes for at least some of
the generated tarballs (see [1]). Since we cannot update the hash
without changing the file name, we switch to Git download in order to
have a different file name.
[1]: https://codeberg.org/Codeberg/Community/issues/2861
Signed-off-by: Raphaël Mélotte <raphael.melotte@mind.be>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Since the bump to version 2.13.0 in Buildroot commit [1],
atop fails to build with musl:
drawbar.c:2336:33: error: implicit declaration of function 'waddwstr'; did you mean 'waddstr'? [-Wimplicit-function-declaration]
(Before 2.13.0, the same configurations already failed earlier on
uClibc because _POSIX_C_SOURCE was defined without a value. Upstream
fixed that in 2.13.0 by giving it an explicit value, but the same
pattern remains for _XOPEN_SOURCE.)
ncurses only declares its wide-character API, including waddwstr(),
when NCURSES_WIDECHAR is set. If the application does not set it,
curses.h derives it from the feature test macros:
#if defined(_XOPEN_SOURCE_EXTENDED) || \
(defined(_XOPEN_SOURCE) && (_XOPEN_SOURCE - 0 >= 500))
drawbar.c defines _XOPEN_SOURCE with an empty value, which makes the
check evaluate as 0 >= 500 and fail.
This only breaks with musl because of the C library's <features.h>,
which curses.h pulls in through <stdio.h> before doing the check.
glibc and uClibc both redefine _XOPEN_SOURCE to 700 and define
_XOPEN_SOURCE_EXTENDED whenever _GNU_SOURCE is set, which drawbar.c
does, so the empty definition gets fixed up. musl never modifies the
application's feature test macros, so the empty value reaches
curses.h and the wide-character prototypes are left out.
Add a patch that defines _XOPEN_SOURCE as 700, matching the
_POSIX_C_SOURCE value of 200809L that is already used.
Fixes:
- https://autobuild.buildroot.org/results/914c1d8fa0fc8bbef64e4a64d58e5941d0077a3e
- https://autobuild.buildroot.org/results/b91ea0722202de0d1fdeeaeea167fd1b21b6d5d6
[1] 2110131cab
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Yegor Yefremov <yegorslists@googlemail.com>
Tested-By: Thomas Perale <thomas.perale@mind.be>
[Julien: add link to Buildroot commit introducing the issue]
Signed-off-by: Julien Olivain <ju.o@free.fr>
The last package depending on libplatform, libcec, removed its
dependency:
332bca975a
The maintainer of both libcec and libplatform has put libplatform into
maintenance mode: "Don't use this library in new projects."
97332651dd
so we can safely remove the package from buildroot.
Signed-off-by: Bernd Kuhls <bernd@kuhls.net>
[Peter: move the COnfig.in.legacy entry under 2026.11]
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Add support for pzstd, the multi-threaded implementation of zstd.
The pzstd build requires C++ and threads support. On some architectures
(e.g., Microblaze, SPARC), it also requires linking against libatomic
due to the use of atomic operations.
Ensure that -latomic is passed in LDFLAGS during both the build and
install phases. Consistently passing LDFLAGS is necessary because the
pzstd Makefile re-links the binary during the 'install' target due to
its dependency on the static library.
Fixes build failures reported with bootlin-microblazeel-uclibc and
bootlin-sparc-uclibc toolchains:
Pzstd.cpp:(.text+0x17e8): undefined reference to '__atomic_compare_exchange_1'
Pzstd.cpp:(.text._ZN5pzstd15BufferWorkQueue3popERNS_6BufferE): undefined reference to '__atomic_fetch_sub_4'
For host-zstd, always build pzstd. It just adds a few seconds of
build time.
Signed-off-by: Shubham Chakraborty <chakrabortyshubham66@gmail.com>
Reviewed-by: Joseph Kogut <joseph.kogut@gmail.com>
Tested-by: Joseph Kogut <joseph.kogut@gmail.com>
[Arnout: unconditionally build pzstd for the host]
Signed-off-by: Arnout Vandecappelle <arnout@rnout.be>
This was added in commit 2d4fd5d358 ("pkg-infra: also set PKGDIR for the
download step") for the DOWNLOAD macro, which was reworked in commit
155f752d8a ("core/download: do not hard-code the path to the package hash
file") to instead use <FOO>_HASH_FILE, so drop it.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
This was added in commit 604d261fb0 ("infra: remove variable
FOO_DIR_PREFIX") to figure out where to find patches, but this was reworked
in commit 99d4180eda ("package/pkg-generic.mk: factor out patch
directories") to instead use <FOO>_PKGDIR, so drop it.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
This was added in commit 7a41cbf938 ("package: fix patching procedure for
host packages") to figure out where to find patches, but this was reworked
in commit 99d4180eda ("package/pkg-generic.mk: factor out patch
directories") to instead use <FOO>_RAWNAME, so drop it.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Since the bump to version 1.9.15 in commit
2a28d25d0c ("package/haveged: bump to version 1.9.15"), haveged
unconditionally includes <semaphore.h> and uses named semaphores
(sem_open() and friends) to synchronize command instances. As a
result, the build fails with toolchains without threads support:
haveged.c:37:10: fatal error: semaphore.h: No such file or directory
Add a dependency on BR2_TOOLCHAIN_HAS_THREADS. Since threads are now
always available, unconditionally pass --enable-threads.
Fixes:
https://autobuild.buildroot.org/results/814cb7b5537d8b7b2c23f3d90e37b2b41de7ee3f
Fixes: 2a28d25d0c ("package/haveged: bump to version 1.9.15")
Signed-off-by: Yegor Yefremov <yegorslists@googlemail.com>
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
We usually put the metadata variables at the beginning of the file and
the variables related to how to build it (like _CONF_OPTS) later. Apply
this principle.
Signed-off-by: Arnout Vandecappelle <arnout@rnout.be>
Fixes:
https://autobuild.buildroot.org/results/e6159cee14f84569c85a61f77824b2b58d68a4bc/
Libtool 2.6.x is starting to be used by upstreams when generating
release tarballs (e.g. cryptsetup-2.8.8). As LIBTOOL_PATCH_HOOK does
not know about it, ltmain.sh is left unpatched, so libtool relinks
libraries at install time and adds -L/usr/lib, which is rejected by
the toolchain wrapper:
i686-buildroot-linux-gnu-gcc: ERROR: unsafe header/library path used in cross-compilation: '-L/usr/lib'
The v2.4.4 patch no longer applies cleanly, so add a v2.6 variant,
rebased on libtool 2.6.2, and use it for 2.6.x versions.
Signed-off-by: Yegor Yefremov <yegorslists@googlemail.com>
Assisted-by: Claude:claude-opus-5
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Commit 8b651fd817 added a dependency on
gcc >= 4.9 and as part of that updated the Config.in comment to
mention C++11 instead of C++. This is not consistent with Buildroot
convention, where we only mention C++, and separately the GCC version
requirement. This commit aligns the Config.in comment with the
Buildroot convention.
Fixes: 8b651fd817 ("package/sqlitecpp: require C++11")
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
SQLiteCpp has an internal setting for stack protection which defaults
to "on". Since Buildroot handles stack protection itself this is not
needed and more importantly it causes problems as the option enables
stack protection support even when the toolchain doesn't support it,
as can be reproduced using the following defconfig:
BR2_TOOLCHAIN_EXTERNAL=y
BR2_TOOLCHAIN_EXTERNAL_BOOTLIN_X86_I686_MUSL_BLEEDING_EDGE=y
BR2_PACKAGE_SQLITECPP=y
Fixes:
br-test-pkg/bootlin-x86-i686-musl
https://gitlab.com/heinemml/buildroot/-/jobs/16865965418
Oddly enough, there are no autobuilder failures recorded for this
issue.
Fixes: e74059b39c ("package/sqlitecpp: new package")
Signed-off-by: Michael Nosthoff <buildroot@heine.tech>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
build and run an SQLiteCpp example to check database writes and queries.
Assisted-by: GitHub Copilot:GPT-6 Sol
Signed-off-by: Michael Nosthoff <buildroot@heine.tech>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
It is sometimes useful to be able to do an offline build, to make sure
the build works properly without network access, so this commit adds a
NO_NETWORK environment variable to our docker-run script, which for
example allows to do:
$ ./utils/docker-run make source
$ NO_NETWORK=1 ./utils/docker-run make
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Upstream changelog:
https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/blob/1.58.1/NEWS?ref_type=tags
Changes relevant for Buildroot build:
* Support for dhclient has been removed, drop the conditional block
for BR2_PACKAGE_DHCP_CLIENT=y.
* NetworkManager now installs a systemd generator to manage its initrd
units, disable it if building without systemd or for systemd without
initrd support by setting the target system generator directory to
"no" [1]. Note that the current upstream main branch adds a build
option for initrd support, likely to be released with 1.60.x, which
should supersede removing the initrd units and disabling the
generator when building with systemd.
* Contrary to what the NEWS file states, the new "clat" build option
is enabled by default. Disable it, (optionally) enabling would
require libbpf and a suitable BPF compiler.
[1] https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/blob/1.58.1/meson.build?ref_type=tags#L416-417
Signed-off-by: Fiona Klute <fiona.klute@gmx.de>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Buildroot boards use a mixture of the GPT Partition Type "L" (generic
Linux filesystem) and "root-x86", "root-arm", etc (UAPI Root
Filesystem) for root filesystem partitions.
To have consistent guidance on building disk images with genimage,
update the docs to recommend using the UAPI Root Filesystem types. The
UAPI Root Filesystem types are part of the UAPI Discoverable Partitions
Specification (DPS) which helps set a wider standard for system boot.
See https://uapi-group.org/specifications/specs/discoverable_partitions_specification/
Signed-off-by: Brandon Maier <brandon.maier@gmail.com>
[Arnout: some additional rewording mainly based on input from Niklas
Cassel]
Signed-off-by: Arnout Vandecappelle <arnout@rnout.be>
Switch to the hatchling build backend.
Add the host-python-hatch-vcs build dependency.
Update license to 0BSD:
64eac3b182
Signed-off-by: James Hilliard <james.hilliard1@gmail.com>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Replace the host-python-mypy dependency with host-python-cython and
enable the Cython extensions instead of mypyc.
Drop the dependency check override now that the Cython requirement is
satisfied.
Signed-off-by: James Hilliard <james.hilliard1@gmail.com>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Switch to the hatchling build backend.
Add the python-anyio runtime dependency.
Add the python-typing-extensions runtime dependency.
Drop the no longer required python-aiofiles dependency.
Signed-off-by: James Hilliard <james.hilliard1@gmail.com>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Switch to the setuptools build backend.
Update license hash due to the copyright year and license URL update:
88f84ed8ad
Signed-off-by: James Hilliard <james.hilliard1@gmail.com>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Update license hash due to copyright year update:
170fccec8c
Add the python-lark runtime dependency.
Drop the no longer required python-ply dependency.
Signed-off-by: James Hilliard <james.hilliard1@gmail.com>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Switch to the flit build backend.
Use the standalone license file, including the updated notices:
42e8f38a90
Signed-off-by: James Hilliard <james.hilliard1@gmail.com>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Since commit 5fc10e76fa2a9 in linux v6.5-rc1, the kernel does a
recursive make call to add the --no-print-directory flag if it is not
set.
Buildroot sets MAKEOVERRIDES empty, which causes none of the
command-line overrides to be propagated to sub-makes. This only goes one
level deep, so it is not a problem. However, if the user sets
MAKEOVERRIDES in their environment, the environment _does_ get
propagated down to sub-makes. Therefore, the recursive call from within
the kernel will not get any of the command-line overrides we give it.
As a result, INSTALL_HDR_PATH will not get passed down to the kernel's
sub-make, thereby breaking the installation to staging.
Note that if Buildroot itself does a sub-make call (e.g. because O= is
set, or umask is different), --no-print-directory is already added so
the issue doesn't occur.
Set --no-print-directory explicitly, which avoids the kernel's sub-make
call, which preserves INSTALL_HDR_PATH. It is only really needed for the
staging install, since the local install in _CONFIGURE_CMDS just sets it
to the default value. For consistency though, set it there as well.
Setting --no-print-directory when it is not needed doesn't hurt.
Signed-off-by: Francesco Mancino <francesco.mancino@tutus.se>
Signed-off-by: Arnout Vandecappelle <arnout@rnout.be>
