Python 3 interprets string literals as Unicode strings, and therefore
\s is treated as an escaped Unicode character which is not correct.
Declaring the RegEx pattern as a raw string instead of unicode is
required for Python 3.
(From OE-Core rev: d4e11eebdfe50acc124a87341721a12bc1c15024)
Signed-off-by: Adrian Freihofer <adrian.freihofer@siemens.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Cherry-picked from master: f2d80817baea298b953d6e14daad65087b3b50c9
Signed-off-by: Adrian Freihofer <adrian.freihofer@siemens.com>
Signed-off-by: Steve Sakoman <steve@sakoman.com>
With recent kernel changes in stable, the full path to python was being
encoded into binaries and python install files. Add some workarounds
for both issues to fix build reproducibility issues.
(From OE-Core rev: 35dbd02b73a586db9e09c6a6d36252244ea5e91f)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
(cherry picked from commit cf33e2ab2693eafaf0e5fc45666355d60e084281)
Signed-off-by: Steve Sakoman <steve@sakoman.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
In 5.19 onwards the build process changed and encoded full build paths
into the output. Adapt the code to look more like our setuptools class
calls. This seems to work ok with older kernels too.
(From OE-Core rev: 6d1bedab94e5ffa1b6fefcf181b09fcf4127bdc7)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
(cherry picked from commit 138673f833a72c636a7fa185089f25dda350dc54)
Signed-off-by: Steve Sakoman <steve@sakoman.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
If you build perf from Linux 5.4 it still contains buildpaths (pointing
to the sysroot's Python) as the existing fixes don't apply to the lines
in that old release.
Add further expressions to remove the final buildpaths.
(From OE-Core rev: a4326da742c41f69c2d1d49cc8fdc2fda74aaa0d)
Signed-off-by: Ross Burton <ross.burton@arm.com>
Signed-off-by: Alexandre Belloni <alexandre.belloni@bootlin.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
(cherry picked from commit 6783c1e295e5afafd73a02fd986ca8956a89dda7)
Signed-off-by: Steve Sakoman <steve@sakoman.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
kernels with upstream commit 630af16eee495f583db5202c3613d1b191f10694
[perf tools: Use Python devtools for version autodetection rather than
runtime], or -stable backports of that commmit, evade our substitutions
for reproducibility.
We add a second sed expression to ensure that our definition of python
is used, as we have the proper environement setup and don't need the
full path of python, which eventually will be captured by the perf
binaries.
(From OE-Core rev: 171415e38e526033a0423f4dc39e9d8e9dc4e5f6)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
(cherry picked from commit b24e73fb34049061ea03a6f2b6a54cdbee7b406b)
Signed-off-by: Steve Sakoman <steve@sakoman.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Commit e1382583cd50 ("perf: sort-pmuevents: don't drop elements") tried
to fix a case where the array terminator elements were dropped from the
sorted list breaking the build, but it only worked for the case where
the terminator is the only element of the array. When the array has other
elements the terminator will still be silently dropped, causing invalid
memory accesses at runtime when the perf utility iterates over the array.
Fix this by treating any unmatched entry as an array terminator and also
add a comment to make it a little more clear how things are ending up at
the right position in the sorted list.
(From OE-Core rev: 70d4a09c1f9fada1a02cf7b3886ffaf39d1b9baf)
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Signed-off-by: Luca Ceresoli <luca.ceresoli@bootlin.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
(cherry picked from commit 69c35a48c5100b884f1b633142b07222b9390e92)
Signed-off-by: Steve Sakoman <steve@sakoman.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
An automated conversion using scripts/contrib/convert-spdx-licenses.py to
convert to use the standard SPDX license identifiers. Two recipes in meta-selftest
were not converted as they're that way specifically for testing. A change in
linux-firmware was also skipped and may need a more manual tweak.
(From OE-Core rev: ceda3238cdbf1beb216ae9ddb242470d5dfc25e0)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
With the changes to binutils, perf's direct calls to ld break for mips n64.
We already have tweaks for n32 so match those with the ones for n64.
(From OE-Core rev: 33828620872b80e1c745612e025215666ba755c8)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
With kernel 5.16 some structs in pmu-events do get a const qualifier, some
a static const storage class and qualifier.
The current sort-pmuevents cannot cope with that and drops all struct
arrays with such additional elements. This then leads to compiler errors.
Allow '^struct', '^const struct', '^static struct', '^static const struct'.
(From OE-Core rev: 8406e83ade1c34d8a7d8063f2e7445aafa471721)
Signed-off-by: Max Krummenacher <max.krummenacher@toradex.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
If a struct element neither has an element cpuid or name it gets silenty
dropped.
Kernel 5.15 for some ARCHs have at least one array of structs matching
this.
e.g. for arm pmu-events.c:
|#include "pmu-events/pmu-events.h"
struct pmu_events_map pmu_events_map[] = {
{
.cpuid = 0,
.version = 0,
.type = 0,
.table = 0,
},
};
struct pmu_sys_events pmu_sys_event_tables[] = {
{
.table = 0
},
};
Before this patch the second struct is translated to an empty array::
struct pmu_sys_events pmu_sys_event_tables[] = {
};
(From OE-Core rev: e1382583cd5060be301afaa9998ccf250cc8bca3)
Signed-off-by: Max Krummenacher <max.krummenacher@toradex.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Those were missed in previous rounds of automated and manual conversion.
(From OE-Core rev: 22f9c7268b542baf6cd8aa0e34c8fb7aa1579e08)
Signed-off-by: Denys Dmytriyenko <denis@denix.org>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
This is the result of automated script conversion:
scripts/contrib/convert-overrides.py <oe-core directory>
converting the metadata to use ":" as the override character instead of "_".
(From OE-Core rev: 42344347be29f0997cc2f7636d9603b1fe1875ae)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
We've seen some reproducibility issues on the autobuilder in perf where the
size of the python module varies slightly between systems. After some head
scratching and removing the --quiet option to the python module build,
you can see it using -Lrecipe-sysroot-native in the linking commandline
for the module. This means it is linking against the native library
on systems where that works, skipping it and using the target one
otherwise, probably with warnings in logs we've not seen.
The fix is to inherit the python3targetconfig class which ensures
that the target sysroot is used, then the byte differences between
the builds go away and things are sane(r) again.
(From OE-Core rev: 8e2b6c042edd9ec76cb8281247604e4f81518780)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
When doing the perf python3 conversion, the audit-python RDEPENDS
was caught up in the regex replacement and was incorrectly changed.
The audit recipe continues to produce a package called audit-python
and it is that package we should have as a RDEPENDS.
(From OE-Core rev: 7eccb9c0c2ea00685451c44cb8faa96c4a2272fd)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Upstream commit:
commit 1dc481c0b0cf18d3952d93a73c4ece90dec277f0
Author: Leo Yan <leo.yan@linaro.org>
Date: Sat Mar 20 18:45:54 2021 +0800
perf test: Change to use bash for daemon test
When executing the daemon test on Arm64 and x86 with Debian (Buster)
distro, both skip the test case with the log:
Changes tools/perf/tests/shell/daemon.sh to be explicitly bash
(it was already required, but was just skipped on various
distros).
We add it into our RDEPENDS for perf-tests to fixup 5.12+
builds.
We already have relatively heavy RDEPENDS for perf tests (python3), so
adding bash into the RDEPENDS isn't signifcant even for older perf
builds that use the same recipe.
(From OE-Core rev: 159cdb159ad0e9d3ed73cfc07f9acd5c0b608e7b)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
perf generates pmu-events.c as part of the build process. The
code that generates the events is doing tree walks and potentially
other non-determinstic things.
We'd rather not mess with that implementation, so we add a script
that knows how to read the pmu-events.c, sort the entries and then
copy it over the generated one.
With this, we should always have events in the same order, improving
reproducibility.
(From OE-Core rev: 5281b2a6e16b6d24b66172b8269478356c0ce6c9)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
perf has been failing our reproducible testing due to multiple symbols
containg build paths.
With this commit, we fix the issues:
1) The following line in the Makefle:
override PYTHON := $(call get-executable-or-default,PYTHON,$(PYTHON_AUTO))
"PYTHON" / "PYTHON_AUTO" have the full path as part of the variable. We've
ensure that the environment is setup and we do not need the full path to be
captured, since the symbol gets built into the executable, making it not
reproducible.
2) The following line:
srcdir_SQ = $(patsubst %tools/perf,tools/perf,$(subst ','\'',$(srcdir))),
Captures the full src path of perf, which of course makes it not
reproducible. We really only need the relative location 'tools/perf', so we
change the Makefile line to remove everything before 'tools/perf'
3) OUTPUT is the full path, we have python on the path so we remove it from the
definition. This is captured in the perf binary, so breaks reproducibility
PYTHONPATH="BUILD_STR(python)
4) To avoid bison generating #ifdefs that have captured paths, we make sure
all the calls have YFLAGS, which contains prefix mapping information.
Upstream-status: OE specific to our cross/build environments. Variants
will be developed for upstream
(From OE-Core rev: dea57aa1cd879df4feae1b9595b38c5b6ff74574)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
* add V=1 to make log.do_compile a bit more useful, with this it shows:
/bin/sh '/OE/build/oe-core/tmp-glibc/work/qemux86_64-oe-linux/perf/1.0-r9/perf-1.0/tools/perf/trace/beauty/arch_errno_names.sh' x86_64-oe-linux-gcc -m64 -march=core2 -mtune=core2 -msse3 -mfpmath=sse --sysroot=/OE/build/oe-core/tmp-glibc/work/qemux86_64-oe-linux/perf/1.0-r9/recipe-sysroot /OE/build/oe-core/tmp-glibc/work/qemux86_64-oe-linux/perf/1.0-r9/perf-1.0/tools > /OE/build/oe-core/tmp-glibc/work/qemux86_64-oe-linux/perf/1.0-r9/perf-1.0/trace/beauty/generated/arch_errno_name_array.c
find: unknown predicate `-m64/arch'
x86_64-oe-linux-gcc: error: unrecognized command-line option '-m64/include/uapi/asm-generic/errno.h'
x86_64-oe-linux-gcc: fatal error: no input files
compilation terminated.
x86_64-oe-linux-gcc: error: unrecognized command-line option '-m64/include/uapi/asm-generic/errno.h'
x86_64-oe-linux-gcc: fatal error: no input files
compilation terminated.
* apply e4ffd066ff
with sed call to fix this
(From OE-Core rev: c603e6e13e44ae907cf1f04abf6e1babf3d7be26)
Signed-off-by: Martin Jansa <Martin.Jansa@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
meta-arm has a recipe for OpenCSD, the CoreSight Trace Decoder library,
so add a PACKAGECONFIG so that users can enable the use of CoreSight in
perf builds.
(From OE-Core rev: 2e9c2f054730069554c9102333aa779c214d8b90)
Signed-off-by: Ross Burton <ross.burton@arm.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Fix error like:
Makefile.perf:597: recipe for target
'intel_x86_64-wrs-linux/perf/1.0-r9/perf-1.0/perf-in.o] Error 2
Makefile.perf:220: recipe for target 'sub-make' failed
make[1]: *** [sub-make] Error 2
Makefile:69: recipe for target 'all' failed
make: *** [all] Error 2
during rebuilding, when kernel version changed, some
source files also maybe changed. It could happened
that the source under ${S} messed up with some already
dropped files and also new files. Fix by clean the ${S}
before copy source files.
(From OE-Core rev: 57b4745cca8d486c1f20b8f01f4a25cb59c4b968)
Signed-off-by: Changqing Li <changqing.li@windriver.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
When the externalsrc class is used the tasks listed in
SRCTREECOVEREDTASKS are deleted to prevent them being executed. If
externalsrc is used for the kernel then this will include
virtual/kernel:do_patch.
We can depend on do_shared_workdir instead as this will survive when
externalsrc is used.
(From OE-Core rev: eab605bab94552046ec4adae5debe026cc03bb4c)
Signed-off-by: Paul Barker <pbarker@konsulko.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
To make the native python3 always used,
- Use sed one-liner instead
- Add substitution for ${S}/scripts/bpf_helpers_doc.py to fix the
following warning.
File "/usr/lib/python3.6/sysconfig.py", line 421, in _init_posix
_temp = __import__(name, globals(), locals(), ['build_time_vars'], 0)
ModuleNotFoundError: No module named '_sysconfigdata'
This issue is first reported by Joel Stanley <joel@jms.id.au>
The sed one-liner is credited to Anuj Mittal <anuj.mittal@intel.com>
(From OE-Core rev: 3f93173130a94310255389cfc62c67102a4fb21b)
Signed-off-by: He Zhe <zhe.he@windriver.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
python code underneath is smart and pokes at python installation in
sysroot for compile environment, the overrides from EXTRA_OEMAKE are
ofcourse preferred but it falls back to python3's distutils/sysconfig
for rest of them, and it does use CCLD and LDSHARED for linking, when we
use clang to compile python3 then it encodes these variables to be clang
however, when building perf we inherit kernel which forces gcc to be
used for compiling perf, which then mixes both compilers due to CCLD and
LDSHARED demands. Therefore override CCLD and LDSHARED in make
environment helps python/distutils to use right linker
(From OE-Core rev: f9259da409c2aa92deb71aef0da2405c8d836615)
Signed-off-by: Khem Raj <raj.khem@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
In kernel 5.5+ there are python3 scripts that explicitly use
/usr/bin/python3 as the interpreter. That will find the host
python and produce undefined results.
We add that interpreter path to our substitutions to ensure
that our sysroot variant is used.
(From OE-Core rev: 103316d50d4947b3c3500eb5cbc4845702a62d22)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The copy of the kernel's top level include directory is not
required to build perf. We have both the linux-libc-headers and
perf's captured/copied headers for what it requires.
The copy of the kernel's headers is leading us to multiple smaller
fixes to ensure that the various .h files are in sync. We can
remove the copy and all of the sync checks, and perf still builds
and executes correctly.
(From OE-Core rev: 79bfd5515887e5faf315a581d98bda63899760da)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
perf can use libcap when available, so we add a packageconfig
for those that want the support.
(From OE-Core rev: 9a35c61d2218e139aa4420ef25f8cf4a0c4c2cfc)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
If we need to copy a single file from the linux source tree into
the perf source tree, we need to ensure that the directory structure
it requires has been created first.
(From OE-Core rev: 7c33f0b1ee35031aa97e5e5231f53d8502b583c9)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
When building perf for 5.4+, we have some new files that need to
be copied (and synchronized) due to structural changes in the
kernel source tree.
Some of the issues these fixes are warnings, but none the less,
they are worth fixing.
- We copy arch/${ARCH}/Makefile, since it is source by some perf
Makefiles
- We copy scripts/, since the perf utilities are looking for files
in that directory stucture.
- We have *three* copies of ctypes.h in the tools/* hierarchy
during the build. If the tools/perf/util/include/linux/ variant
is used, it will trigger build errors since it is not complete.
We copy the kernel's main include/linux/ctype.h to ensure they
are in sync.
(From OE-Core rev: 6a000d734c2fe57b4a601fe7e7373096345b02b4)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
* tools/include/linux/bits.h was added in v4.20-rc1 with this commit:
commit ba4aa02b417f08a0bee5e7b8ed70cac788a7c854
Author: Arnaldo Carvalho de Melo <acme@redhat.com>
Date: Tue Sep 25 10:55:59 2018 -0300
tools include: Adopt linux/bits.h
* also if you're building for such older kernel you will probably see
do_compile failing with:
| config/Makefile:448: Missing perl devel files. Disabling perl scripting support, please install perl-ExtUtils-Embed/libperl-dev
| config/Makefile:495: Python 3 is not yet supported; please set
| config/Makefile:496: PYTHON and/or PYTHON_CONFIG appropriately.
easiest work around is to disable scripting PACKAGECONFIG, because
since oe-core commit:
commit 584af667e0000129bcb5c9e8108485f2f6590eaf
Author: Bruce Ashfield <bruce.ashfield@gmail.com>
Date: Wed Aug 28 22:14:41 2019 -0400
perf: change dependencies on python to python3
The upstream kernel can now handle python3 for the perf scripts, coupled
with the impending EOL of python2, we switch the dependencies in perf
(scripting) to python3.
it now uses python3, but the support for that was added in kernel
v4.17-rc1 with:
commit 66dfdff03d196e51322c6a85c0d8db8bb2bdd655
Author: Jaroslav Skarvada <jskarvad@redhat.com>
Date: Fri Jan 19 21:56:41 2018 +0100
perf tools: Add Python 3 support
Added Python 3 support while keeping Python 2.7 compatibility.
if you really need scripting support than either backport the kernel
patch to your kernel or undo the perf recipe changes.
(From OE-Core rev: d6ee3d3bdf183a9060b5b0c25b97688a4b9d40dc)
Signed-off-by: Martin Jansa <Martin.Jansa@gmail.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
In the 5.3+ perf builds, there are multiple unistd.h files that need
to be kept in sync. So not only do we update one, we update both to
ensure that libc-headers and the perf build are in sync.
perf checks some key files itself to be sure they are in sync, and
bits.h can be out of sync. There's no build error that results with
bits.h at the moment, but we still make sure they are in sync to
avoid the configure/compile warning.
(From OE-Core rev: b850ee597e7116cc2a7546853f2b976c891196ef)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The upstream kernel can now handle python3 for the perf scripts, coupled
with the impending EOL of python2, we switch the dependencies in perf
(scripting) to python3.
Outside of dependency changes, we also pass the following two variables
to the perf build: PYTHON=python3 PYTHON_CONFIG=python3-config. To
ensure that our python3 components are used.
For configurations that disable perf scripting, there is no change in
behaviour.
[YOCTO #13358]
(From OE-Core rev: 584af667e0000129bcb5c9e8108485f2f6590eaf)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
* before trying to copy unistd.h into it
* older kernels don't have uapi in tools/include and do_configure fails
with:
DEBUG: Executing shell function do_configure
cp: cannot create regular file '.../perf/1.0-r9/perf-1.0/tools/include/uapi/asm-generic/unistd.h':
No such file or directory
WARNING: exit code 1 from a shell command.
* tools/include/uapi was added in kernel 4.8 with
c4b6014e8b
tools: Add copy of perf_event.h to tools/include/linux/
(From OE-Core rev: 5fdb96f4996856286c4f72b93d31b7990ceed8aa)
Signed-off-by: Martin Jansa <Martin.Jansa@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
When enable DEBUG_BUILD, the perf build failed by the following error:
libbpf.c:727:36: error: 'data' may be used uninitialized in this function [-Werror=maybe-uninitialized]
This is ok until Khem commit a patch in oe-core:
16643b03227466e2c80a24c2d079fe36e89553c1
This commit import "-Og" option to "DEBUG_OPTIMIZATION", result in this error.
Actually, the warning was misinformation. We explore the code in libbpf.c:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/bpf/libbpf.c#n640
--------------------------------------------------------------------------
if (scn)
data = elf_getdata(scn, NULL);
if (!scn || !data) {
pr_warning("failed to get Elf_Data from map section %d\n",
obj->efile.maps_shndx);
return -EINVAL;
}
......
using data...
......
--------------------------------------------------------------------------
If the 'scn' was not null, the data will be initialized. if the data was null,
the function will return -EINVAL, so the data must be not null, but compiler
can't see it in 'if' code block, so treat it as 'maybe-uninitialized'.
(From OE-Core rev: 0eb4048bc5865579327f5a88086e3e3241b012cb)
Signed-off-by: Dengke Du <dengke.du@windriver.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
During the build for some architectures, perf generates a program
which executes on the host to dump the syscall table.
The generation of that program uses the cross compiler + sysroot
to expand unistd.h. As such, we are getting the contents of that
file from linux-libc-headers.
The compilation of that generated program uses the host compiler
and a restricted include path to the perf source code. In the
perf source there is a captured unistd.h, as such it will be used
when compiling the host executable. The perf source code is copied
from the kernel version that is being built .. so we have a
mismatch between the generation and the compilation of the host
program.
Normally this mismatch is fine, but if the libc-headers are
newer than the kernel, we'll have syscalls (and their syscall
numbers) that are not defined in the perf source code. This
leads to a compiler error and a cascading failure of the perf
build due to a missing generated file.
To fix this, we can copy unistd.h from the recipe-sysroot
into the perf source code and they will always be in sync.
(From OE-Core rev: 3a23e094c4189af878e467f1636aea63955e250d)
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
This helps it find the unwind library in sysroot always and not
look around and confuse with host
(From OE-Core rev: e914c03964c540493523b18660f8a4eaeec78e3d)
Signed-off-by: Khem Raj <raj.khem@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
perf's jvmti support builds libperf-jvmti.so that needs to be packaged
as part of the main package, not a developer package. Given there's no
files in the developer package, override FILES_SOLIBSDEV so it's not
gobbled by FILES_${PN}-dev, and then add just that file to FILES_${PN}
so that if in the future there other files which would be added to the
dev package we'll fail then.
(From OE-Core rev: 284763fcf188a497f15a1c69a50ceeee99d546f8)
Signed-off-by: Alex Kiernan <alex.kiernan@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
libunwind is not yet ported for ARC & RISCV64 and on attempt
to build it for those arches we just get an error message.
If we explicitly disable libunwind it is gracefully handled by
perf build system and it just gets configured to not use it
so perf is still usable even on those arches.
(From OE-Core rev: bb06900a1f29fdd4066f6d7e7a961d230c2e9438)
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Cc: Khem Raj <raj.khem@gmail.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
On a builder which has /tmp mounted nexec, the ARM64 generation of
the sycall table will fail:
[snip]
|perf-1.0/tools/perf/arch/arm64/entry/syscalls//mksyscalltbl:
/tmp/create-table-6VGPSt: Permission denied
[snip]
To avoid the permission denied error, we must instruct the perf build
to place temporary files in an appropriate directory (versus always
using /tmp). To do this, we need both a kernel source change to perf,
and a recipe change to pass the temp directory in by a TMPDIR variable.
The linux-yocto has a reference kernel change: [perf arm64: Fix generate
system call table failed with /tmp mounted with noexec], that has also
been submitted upstream mailing list linux-perf-users@vger.kernel.org
(From OE-Core rev: d819e7fa76e2b732aa7c33ab0e9a834781090824)
Signed-off-by: Hongxu Jia <hongxu.jia@windriver.com>
Reviewed-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
A missing src directory from a broken kernel recipe resulted
only in a warning:
WARNING: copyfile: stat of /home/builder/src/tmp-glibc/work-shared/target/kernel-source/tools/arch failed ([Errno 2] No such file or directory: '/home/builder/src/tmp-glibc/work-shared/target/kernel-source/tools/arch')
and the build failed horribly.
With this change it's an error which can not be missed:
ERROR: perf-1.0-r9 do_configure: Path does not exist: /home/builder/src/tmp-glibc/work-shared/target/kernel-source/tools/arch. Maybe PERF_SRC does not match the kernel version.
ERROR: perf-1.0-r9 do_configure: Function failed: copy_perf_source_from_kernel
ERROR: Logfile of failure stored in: /home/builder/src/tmp-glibc/work/target-linux/perf/1.0-r9/temp/log.do_configure.21083
NOTE: recipe perf-1.0-r9: task do_configure: Failed
ERROR: Task (/home/builder/src/poky/meta/recipes-kernel/perf/perf.bb:do_configure) failed with exit code '1'
To get get perf compiling from a custom kernel, a perf.bbappend can be
created which defines PERF_SRC as list of files and directories needed
from kernel source tree to compile perf. This varies between kernel
versions.
(From OE-Core rev: c392dabefc431dbfb31d6a1465c75ba9cc765804)
Signed-off-by: Mikko Rapeli <mikko.rapeli@bmw.de>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Whenever perf got rebuilt, I was consistently getting errors such as
| find: '[...]/perf/1.0-r9/perf-1.0/plugin_mac80211.so': No such file or directory
| find: '[...]/perf/1.0-r9/perf-1.0/plugin_mac80211.so': No such file or directory
| find: find: '[...]/perf/1.0-r9/perf-1.0/libtraceevent.a''[...]/perf/1.0-r9/perf-1.0/libtraceevent.a': No such file or directory: No such file or directory
|
[...]
| find: cannot delete '/mnt/xfs/devel/pil/yocto/tmp-glibc/work/wandboard-oe-linux-gnueabi/perf/1.0-r9/perf-1.0/util/.pstack.o.cmd': No such file or directory
breaking the whole build. The root cause seems to be that the implicit
'make clean' done during do_configure ends up running in parallel, and
thus multiple find commands attempt to stat and/or delete the same
file.
A patch disabling parallelism for the clean target has been ack'ed
upstream (lkml.kernel.org/r/20180705134955.GB3686@krava), but it should
be harmless to pass JOBS=1 even with a fixed kernel. This can be removed
if and when all relevant -stable kernels have that patch.
(From OE-Core rev: bb58203b668df42fd08c2e5fa4a172cf63e37369)
Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
* perf is failing to build for me since this oe-core commit:
commit 9b38c824961fc9dce51bda95c25dac91a69fc64f
Author: Hongxu Jia <hongxu.jia@windriver.com>
Date: Tue Apr 24 11:33:47 2018 +0800
perf: make a copy of kernel source to perf workdir
the problem is that perf sources in kernel older than 4.8 (in my case
4.4) are depending on the "global" include headers outside tools
directory, e.g. swab.h in:
kernel-source/tools$ git grep swab.h
perf/MANIFEST:include/linux/swab.h
perf/MANIFEST:include/uapi/linux/swab.h
perf/util/include/asm/byteorder.h:#include "../../../../include/uapi/linux/swab.h"
this was resolved in 4.8 with:
commit 7e3f36411342a54f1981fa97b43550b8406a3d69
Author: Arnaldo Carvalho de Melo <acme@redhat.com>
Date: Mon Jul 18 17:42:16 2016 -0300
perf tools: Remove tools/perf/util/include/asm/byteorder.h
Not used anymore. This also stops include linux/swab.h directly
from the kernel sources, remove that reference from the MANIFEST.
and few more changes to make tools/include more complete and standalone:
tools/include in 4.15:
asm asm-generic linux tools trace uapi
tools/include in 4.4:
asm asm-generic linux tools
but copying the include header even for kernels which don't really
need it doesn't add big overhead, so just copy include to perf sources
for all kernels.
(From OE-Core rev: 19fb2d11a8bb3c6dfdd5edc1b9155d642dc0f5e0)
Signed-off-by: Martin Jansa <Martin.Jansa@gmail.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Since perf contaminates linux shared workdir, it probably caused
kernel-devsrc compile failure at world build.
...
|0 blocks
|cpio: ./tools/perf/arch/arm/util/sedr7ORqk: Cannot stat:
No such file or directory
|0 blocks
...
cpio tried to find a file at ${S}/tools/perf and failed
if the input list is not valid.
Make a copy of kernel shared source directory into a perf workdir
could fix the issue.
Drop `Fix for rebuilding' which is obsolete
[YOCTO #10880]
(From OE-Core rev: 9b38c824961fc9dce51bda95c25dac91a69fc64f)
Signed-off-by: Hongxu Jia <hongxu.jia@windriver.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
perf needs audit-python to be able to show syscall names and for
'perf trace' to work.
Enable dependency on audit-python if present in PACKAGECONFIG. It's
disabled by default since audit as of now is in meta-selinux.
Fixes [YOCTO #3343]
Fixes [YOCTO #3358]
(From OE-Core rev: c386abacae89a148e77ffa51630c7917e90406f9)
Signed-off-by: Anuj Mittal <anuj.mittal@intel.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
