linux-kernel - Re: [PATCH 2/2] perf clang: Fixes for more recent LLVM/clang

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CAP-5=fXv7oxrxBnHqXNAdwK4ULPUcAVR5aFh1RV1NsTX3JuYxg@mail.gmail.com>
Date:   Fri, 12 Nov 2021 07:35:49 -0800
From:   Ian Rogers <irogers@...gle.com>
To:     Ma Xinjian <xinjianx.ma@...el.com>
Cc:     acme@...nel.org, alexander.shishkin@...ux.intel.com,
        daniel@...earbox.net, jolsa@...hat.com, leo.yan@...aro.org,
        linux-kernel@...r.kernel.org, linux-perf-users@...r.kernel.org,
        llvm@...ts.linux.dev, mark.rutland@....com, mingo@...hat.com,
        mpetlan@...hat.com, namhyung@...nel.org, nathan@...nel.org,
        ndesaulniers@...gle.com, peterz@...radead.org,
        sedat.dilek@...il.com, philip.li@...el.com
Subject: Re: [PATCH 2/2] perf clang: Fixes for more recent LLVM/clang

On Fri, Nov 12, 2021 at 2:34 AM Ma Xinjian <xinjianx.ma@...el.com> wrote:
>
> Hi
>
> We build perf with clang also recently, but unlucky that failed with
> following errors:
>
>
> ```
> In file included from /usr/lib/llvm-7/include/llvm/ADT/STLExtras.h:21,
>                   from /usr/lib/llvm-7/include/llvm/ADT/StringRef.h:13,
>                   from /usr/lib/llvm-7/include/clang/Basic/Version.h:20,
>                   from test-clang.cpp:2:
> /usr/lib/llvm-7/include/llvm/ADT/SmallVector.h: In instantiation of
> 'void llvm::SmallVectorTemplateBase<T, true>::push_back(const T&) [with
> T = clang::driver::OffloadUnbundlingJobAction::DependentActionInfo]':
> /usr/lib/llvm-7/include/clang/Driver/Action.h:579:61:   required from here
> /usr/lib/llvm-7/include/llvm/ADT/SmallVector.h:313:11: error: 'void*
> memcpy(void*, const void*, size_t)' writing to an object of type 'struct
> clang::driver::OffloadUnbundlingJobAction::DependentActionInfo' with no
> trivial copy-assignment; use copy-initialization instead
> [-Werror=class-memaccess]
>       memcpy(this->end(), &Elt, sizeof(T));
>       ~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> In file included from /usr/lib/llvm-7/include/clang/Driver/Driver.h:15,
>                   from test-clang.cpp:6:
> /usr/lib/llvm-7/include/clang/Driver/Action.h:549:10: note: 'struct
> clang::driver::OffloadUnbundlingJobAction::DependentActionInfo' declared
> here
>     struct DependentActionInfo final {
>            ^~~~~~~~~~~~~~~~~~~
> cc1plus: all warnings being treated as errors
> ```
>
> llvm version: 7
> rootfs: Debian10
> kernel: v5.8 v5.11 v5.15
> reproduce:
> $ cd linux/tools/perf
> $ make LIBCLANGLLVM=1 ARCH= EXTRA_CFLAGS='-fno-omit-frame-pointer
> -fsanitize=undefined -fsanitize=address'
>
> Would you please kindly point out what's wrong with our steps? Or can
> you give us your command to make perf with clang?

Hi Xinjian,

As the bug is in LLVM 7 there is little we can do in Linux perf to fix
the issue. I haven't got LLVM 7 on my machine to test with, but I see
the same problem with LLVM 6 - both of which are of course quite old.
My workaround is to add to the make command line
"CXXFLAGS=-Wno-error=class-memaccess". My full command line is:

make -C tools/perf O=/tmp/perf DESTDIR=/tmp/perf-install  DEBUG=1
LIBPFM4=1 NO_LIBBFD=1 LIBCLANGLLVM=1
LLVM_CONFIG=/usr/lib/llvm-6.0/bin/llvm-config
CXXFLAGS=-Wno-error=class-memaccess

I know Intel is keen on testing. Related to this code are the BPF
tests in "perf test". I frequently see flakes on these tests that
count the number of times a system call is called. The flakes are an
under or over count by 1. The over count by 1 I can explain if during
the test another process happens to use the same system call. The
under counting is harder to explain and is possibly an indication of a
data race. I'd be happy for any suggestions to fix these flakes.

Thanks,
Ian

> Thanks
> Ma Xinjian