Re: Linux-5.4: bpf: test_core_reloc_arrays.o: Segmentation fault with llc -march=bpf

[Yonghong Song <yhs@fb.com>]

On 10/24/19 11:31 AM, Ilya Leoshkevich wrote:
>> Am 24.10.2019 um 19:49 schrieb Yonghong Song <yhs@fb.com>:
>>
>>
>>
>> On 10/24/19 9:04 AM, Ilya Leoshkevich wrote:
>>>> Am 23.10.2019 um 03:35 schrieb Prabhakar Kushwaha <prabhakar.pkin@gmail.com>:
>>>>
>>>>
>>>> Adding other mailing list, folks...
>>>>
>>>> Hi All,
>>>>
>>>> I am trying to build kselftest on Linux-5.4 on ubuntu 18.04. I installed
>>>> LLVM-9.0.0 and Clang-9.0.0 from below links after following steps from
>>>> [1] because of discussion [2]
>>>>
>>>> https://urldefense.proofpoint.com/v2/url?u=https-3A__releases.llvm.org_9.0.0_llvm-2D9.0.0.src.tar.xz&d=DwIFAg&c=5VD0RTtNlTh3ycd41b3MUw&r=DA8e1B5r073vIqRrFz7MRA&m=se8pV6OlDAeF2g5iEAvSB2qhLBJGPaHADv3NQVNFx6U&s=IzBxNhAvcILfAD_XcSB7t0s6-B-wFY3TBoVGH6WhRK8&e=
>>>> https://urldefense.proofpoint.com/v2/url?u=https-3A__releases.llvm.org_9.0.0_clang-2Dtools-2Dextra-2D9.0.0.src.tar.xz&d=DwIFAg&c=5VD0RTtNlTh3ycd41b3MUw&r=DA8e1B5r073vIqRrFz7MRA&m=se8pV6OlDAeF2g5iEAvSB2qhLBJGPaHADv3NQVNFx6U&s=KkjCjWm_q2iMfFh50rTKtFqQEMbRBVhT9Oh8KMfgwW4&e=
>>>> https://urldefense.proofpoint.com/v2/url?u=https-3A__releases.llvm.org_9.0.0_cfe-2D9.0.0.src.tar.xz&d=DwIFAg&c=5VD0RTtNlTh3ycd41b3MUw&r=DA8e1B5r073vIqRrFz7MRA&m=se8pV6OlDAeF2g5iEAvSB2qhLBJGPaHADv3NQVNFx6U&s=TvkN9sb5rSB5BNxJP27UmCsfNHsRQdaVeAnBa1TkyjM&e=
>>>>
>>>> Now, i am trying with llc -march=bpf, with this segmentation fault is
>>>> coming as below:
>>>>
>>>> gcc -g -Wall -O2 -I../../../include/uapi -I../../../lib
>>>> -I../../../lib/bpf -I../../../../include/generated -DHAVE_GENHDR
>>>> -I../../../include -Dbpf_prog_load=bpf_prog_test_load
>>>> -Dbpf_load_program=bpf_test_load_program    test_flow_dissector.c
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/test_stub.o
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/libbpf.a -lcap -lelf
>>>> -lrt -lpthread -o
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/test_flow_dissector
>>>> gcc -g -Wall -O2 -I../../../include/uapi -I../../../lib
>>>> -I../../../lib/bpf -I../../../../include/generated -DHAVE_GENHDR
>>>> -I../../../include -Dbpf_prog_load=bpf_prog_test_load
>>>> -Dbpf_load_program=bpf_test_load_program
>>>> test_tcp_check_syncookie_user.c
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/test_stub.o
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/libbpf.a -lcap -lelf
>>>> -lrt -lpthread -o
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/test_tcp_check_syncookie_user
>>>> gcc -g -Wall -O2 -I../../../include/uapi -I../../../lib
>>>> -I../../../lib/bpf -I../../../../include/generated -DHAVE_GENHDR
>>>> -I../../../include -Dbpf_prog_load=bpf_prog_test_load
>>>> -Dbpf_load_program=bpf_test_load_program    test_lirc_mode2_user.c
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/test_stub.o
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/libbpf.a -lcap -lelf
>>>> -lrt -lpthread -o
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/test_lirc_mode2_user
>>>> (clang -I. -I./include/uapi -I../../../include/uapi
>>>> -I/usr/src/tovards/linux/tools/testing/selftests/bpf/../usr/include
>>>> -D__TARGET_ARCH_arm64 -g -idirafter /usr/local/include -idirafter
>>>> /usr/local/lib/clang/9.0.0/include -idirafter
>>>> /usr/include/aarch64-linux-gnu -idirafter /usr/include
>>>> -Wno-compare-distinct-pointer-types -O2 -target bpf -emit-llvm \
>>>> -c progs/test_core_reloc_arrays.c -o - || echo "clang failed") | \
>>>> llc -march=bpf -mcpu=probe  -filetype=obj -o
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/test_core_reloc_arrays.o
>>>> Stack dump:
>>>> 0. Program arguments: llc -march=bpf -mcpu=probe -filetype=obj -o
>>>> /usr/src/tovards/linux/tools/testing/selftests/bpf/test_core_reloc_arrays.o
>>>> 1. Running pass 'Function Pass Manager' on module '<stdin>'.
>>>> 2. Running pass 'BPF Assembly Printer' on function '@test_core_arrays'
>>>> #0 0x0000aaaac618db08 llvm::sys::PrintStackTrace(llvm::raw_ostream&)
>>>> (/usr/local/bin/llc+0x152eb08)
>>>> Segmentation fault
>>>
>>> Hi,
>>>
>>> FWIW I can confirm that this is happening on s390 too with llvm-project
>>> commit 950b800c451f.
>>>
>>> Here is the reduced sample that triggers this (with -march=bpf
>>> -mattr=+alu32):
>>>
>>> struct b {
>>>    int e;
>>> } c;
>>> int f() {
>>>    return __builtin_preserve_field_info(c.e, 0);
>>> }
>>>
>>> This is compiled into:
>>>
>>> 0B      bb.0 (%ir-block.0):
>>> 16B       %0:gpr = LD_imm64 @"b:0:0$0:0"
>>> 32B       $w0 = COPY %0:gpr, debug-location !17; 1-E.c:5:3
>>> 48B       RET implicit killed $w0, debug-location !17; 1-E.c:5:3
>>>
>>> and then BPFInstrInfo::copyPhysReg chokes on COPY, since $w0 and %0 are
>>> in different register classes.
>>
>> Ilya,
>>
>> Thanks for reporting. I can reproduce the issue with latest trunk.
>> I will investigate and fix the problem soon.
>>
>> Yonghong
> 
> Thanks for taking care of this! Just FYI, bisect pointed to 05e46979d2f4
> ("[BPF] do compile-once run-everywhere relocation for bitfields").
> 
> Could you please add me to Phabricator review? I'm curious what the

I did add you in the diff https://reviews.llvm.org/D69438.

> proper solution is going to be, as I'm still not sure whether handling
> asymmetric copies is the right approach, or whether they should rather
> be prevented from occuring in the first place.

The change has been pushed into the trunk.
We indeed used asymmetric copy. Please do let me know if you think
there is a better way to do that. Thanks!

> 
> Best regards,
> Ilya
>