[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <1b7ae999-c081-45c8-a914-f215c829d57e@linux.dev>
Date: Sat, 1 Feb 2025 22:22:07 -0800
From: Yonghong Song <yonghong.song@...ux.dev>
To: Daniel Xu <dxu@...uu.xyz>
Cc: shuah@...nel.org, andrii@...nel.org, eddyz87@...il.com, ast@...nel.org,
nathan@...nel.org, daniel@...earbox.net, martin.lau@...ux.dev,
song@...nel.org, john.fastabend@...il.com, kpsingh@...nel.org,
sdf@...ichev.me, haoluo@...gle.com, jolsa@...nel.org, mykolal@...com,
ndesaulniers@...gle.com, morbo@...gle.com, justinstitt@...gle.com,
bpf@...r.kernel.org, linux-kselftest@...r.kernel.org,
linux-kernel@...r.kernel.org, llvm@...ts.linux.dev
Subject: Re: [PATCH] selftests: bpf: Support dynamic linking LLVM if static
not available
On 2/1/25 12:23 AM, Daniel Xu wrote:
> Hi Yonghong,
>
> On Thu, Jan 30, 2025 at 10:28:11PM -0800, Yonghong Song wrote:
>>
>>
>> On 1/30/25 2:33 PM, Daniel Xu wrote:
>>> Since 67ab80a01886 ("selftests/bpf: Prefer static linking for LLVM
>>> libraries"), only statically linking test_progs is supported. However,
>>> some distros only provide a dynamically linkable LLVM.
>>>
>>> This commit adds a fallback for dynamically linking LLVM if static
>>> linking is not available. If both options are available, static linking
>>> is chosen.
>>>
>>> Signed-off-by: Daniel Xu <dxu@...uu.xyz>
>>> ---
>>> tools/testing/selftests/bpf/Makefile | 11 ++++++++---
>>> 1 file changed, 8 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile
>>> index 6722080b2107..da514030a153 100644
>>> --- a/tools/testing/selftests/bpf/Makefile
>>> +++ b/tools/testing/selftests/bpf/Makefile
>>> @@ -184,9 +184,14 @@ ifeq ($(feature-llvm),1)
>>> LLVM_CONFIG_LIB_COMPONENTS := mcdisassembler all-targets
>>> # both llvm-config and lib.mk add -D_GNU_SOURCE, which ends up as conflict
>>> LLVM_CFLAGS += $(filter-out -D_GNU_SOURCE,$(shell $(LLVM_CONFIG) --cflags))
>>> - LLVM_LDLIBS += $(shell $(LLVM_CONFIG) --link-static --libs $(LLVM_CONFIG_LIB_COMPONENTS))
>>> - LLVM_LDLIBS += $(shell $(LLVM_CONFIG) --link-static --system-libs $(LLVM_CONFIG_LIB_COMPONENTS))
>>> - LLVM_LDLIBS += -lstdc++
>>> + # Prefer linking statically if it's available, otherwise fallback to shared
>>> + ifeq ($(shell $(LLVM_CONFIG) --link-static --libs &> /dev/null && echo static),static)
>>> + LLVM_LDLIBS += $(shell $(LLVM_CONFIG) --link-static --libs $(LLVM_CONFIG_LIB_COMPONENTS))
>>> + LLVM_LDLIBS += $(shell $(LLVM_CONFIG) --link-static --system-libs $(LLVM_CONFIG_LIB_COMPONENTS))
>>> + LLVM_LDLIBS += -lstdc++
>>> + else
>>> + LLVM_LDLIBS += $(shell $(LLVM_CONFIG) --link-shared --libs $(LLVM_CONFIG_LIB_COMPONENTS))
>>> + endif
>>> LLVM_LDFLAGS += $(shell $(LLVM_CONFIG) --ldflags)
>>> endif
>> Although your change looks good, but maybe you can look at bpftool Makefile?
>>
>> # If LLVM is available, use it for JIT disassembly
>> CFLAGS += -DHAVE_LLVM_SUPPORT
>> LLVM_CONFIG_LIB_COMPONENTS := mcdisassembler all-targets
>> # llvm-config always adds -D_GNU_SOURCE, however, it may already be in CFLAGS
>> # (e.g. when bpftool build is called from selftests build as selftests
>> # Makefile includes lib.mk which sets -D_GNU_SOURCE) which would cause
>> # compilation error due to redefinition. Let's filter it out here.
>> CFLAGS += $(filter-out -D_GNU_SOURCE,$(shell $(LLVM_CONFIG) --cflags))
>> LIBS += $(shell $(LLVM_CONFIG) --libs $(LLVM_CONFIG_LIB_COMPONENTS))
>> ifeq ($(shell $(LLVM_CONFIG) --shared-mode),static)
>> LIBS += $(shell $(LLVM_CONFIG) --system-libs $(LLVM_CONFIG_LIB_COMPONENTS))
>> LIBS += -lstdc++
>> endif
>> LDFLAGS += $(shell $(LLVM_CONFIG) --ldflags)
>>
>> It would be great if the selftests shared library handling to be the same as bpftool's.
> So bpftool is both an internally consumed (from selftests) dependency as
> well as a tool packaged up by distros. For the latter case, distros
> prefer dynamic linking.
I hacked llvm to have both static and shared libraries installed and indeed
`llvm-config --shared-mode` prefers shared mode.
So yes, your existing change looks good. Thanks.
>
> So unfortunately, I think these probably need to be defined separately.
> The code looks similar but the use cases are different.
>
> Thanks,
> Daniel
Powered by blists - more mailing lists