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  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]
Date:   Mon, 16 Mar 2020 17:56:51 -0700
From:   Fangrui Song <maskray@...gle.com>
To:     Stanislav Fomichev <sdf@...ichev.me>
Cc:     Stanislav Fomichev <sdf@...gle.com>, netdev@...r.kernel.org,
        bpf@...r.kernel.org, davem@...emloft.net, ast@...nel.org,
        daniel@...earbox.net, Nick Desaulniers <ndesaulniers@...gle.com>,
        Nathan Chancellor <natechancellor@...il.com>,
        clang-built-linux@...glegroups.com
Subject: Re: [PATCH bpf] bpf: Support llvm-objcopy for vmlinux BTF

On 2020-03-16, Stanislav Fomichev wrote:
>On 03/16, Fangrui Song wrote:
>> On 2020-03-16, Stanislav Fomichev wrote:
>> > Commit da5fb18225b4 ("bpf: Support pre-2.25-binutils objcopy for vmlinux
>> > BTF") switched from --dump-section to
>> > --only-section/--change-section-address for BTF export assuming
>> > those ("legacy") options should cover all objcopy versions.
>> >
>> > Turns out llvm-objcopy doesn't implement --change-section-address [1],
>> > but it does support --dump-section. Let's partially roll back and
>> > try to use --dump-section first and fall back to
>> > --only-section/--change-section-address for the older binutils.
>> >
>> > 1. https://bugs.llvm.org/show_bug.cgi?id=45217
>> >
>> > Fixes: df786c9b9476 ("bpf: Force .BTF section start to zero when dumping from vmlinux")
>> > Tested-by: Nick Desaulniers <ndesaulniers@...gle.com>
>> > Reported-by: Nathan Chancellor <natechancellor@...il.com>
>> > Link: https://github.com/ClangBuiltLinux/linux/issues/871
>> > Signed-off-by: Stanislav Fomichev <sdf@...gle.com>
>> > ---
>> > scripts/link-vmlinux.sh | 10 ++++++++++
>> > 1 file changed, 10 insertions(+)
>> >
>> > diff --git a/scripts/link-vmlinux.sh b/scripts/link-vmlinux.sh
>> > index dd484e92752e..8ddf57cbc439 100755
>> > --- a/scripts/link-vmlinux.sh
>> > +++ b/scripts/link-vmlinux.sh
>> > @@ -127,6 +127,16 @@ gen_btf()
>> > 		cut -d, -f1 | cut -d' ' -f2)
>> > 	bin_format=$(LANG=C ${OBJDUMP} -f ${1} | grep 'file format' | \
>> > 		awk '{print $4}')
>> > +
>> > +	# Compatibility issues:
>> > +	# - pre-2.25 binutils objcopy doesn't support --dump-section
>> > +	# - llvm-objcopy doesn't support --change-section-address, but
>> > +	#   does support --dump-section
>> > +	#
>> > +	# Try to use --dump-section which should cover both recent
>> > +	# binutils and llvm-objcopy and fall back to --only-section
>> > +	# for pre-2.25 binutils.
>> > +	${OBJCOPY} --dump-section .BTF=$bin_file ${1} 2>/dev/null || \
>> > 	${OBJCOPY} --change-section-address .BTF=0 \
>> > 		--set-section-flags .BTF=alloc -O binary \
>> > 		--only-section=.BTF ${1} .btf.vmlinux.bin
>> > --
>> > 2.25.1.481.gfbce0eb801-goog
>>
>> So let me take advantage of this email to ask some questions about
>> commit da5fb18225b4 ("bpf: Support pre-2.25-binutils objcopy for vmlinux BTF").
>>
>> Does .BTF have the SHF_ALLOC flag?
>No, that's why we manually do '--set-section-flags .BTF=alloc' to
>make --only-section work.
>
>> Is it a GNU objcopy<2.25 bug that objcopy --set-section-flags .BTF=alloc -O binary --only-section=.BTF does not skip the content?
>> Non-SHF_ALLOC sections usually have 0 sh_addr. Why do they need --change-section-address .BTF=0 at all?
>I think that '--set-section-flags .BTF=alloc' causes objcopy to put
>some non-zero (valid) sh_addr, that's why we need to reset it to 0.
>
>(it's not clear if it's a feature or a bug and man isn't helpful)
>
>> Regarding
>>
>> > Turns out llvm-objcopy doesn't implement --change-section-address [1],
>>
>> This option will be difficult to implement in llvm-objcopy if we intend
>> it to have a GNU objcopy compatible behavior.
>> Without --only-section, it is not very clear how
>> --change-section-{address,vma,lma} will affect program headers.
>> There will be a debate even if we decide to implement them in llvm-objcopy.
>>
>> Some PT_LOAD rewriting examples:
>>
>>   objcopy --change-section-address .plt=0 a b
>>   objcopy --change-section-address .text=0 a b
>>
>> There is another bug related to -B
>> (https://github.com/ClangBuiltLinux/linux/issues/871#issuecomment-599790909):
>>
>> + objcopy --change-section-address .BTF=0 --set-section-flags .BTF=alloc
>> -O binary --only-section=.BTF .tmp_vmlinux.btf .btf.vmlinux.bin
>> + objcopy -I binary -O elf64-x86-64 -B x86_64 --rename-section .data=.BTF .btf.vmlinux.bin .btf.vmlinux.bin.o
>> objcopy: architecture x86_64 unknown
>> + echo 'Failed to generate BTF for vmlinux'
>>
>> It should be i386:x86_64.
>Here is what I get:
>
>+ bin_arch=i386:x86-64
>+ bin_format=elf64-x86-64
>+ objcopy --change-section-address .BTF=0 --set-section-flags .BTF=alloc -O binary --only-section=.BTF .tmp_vmlinux.btf .btf.vmlinux.bin
>+ objcopy -I binary -O elf64-x86-64 -B i386:x86-64 --rename-section .data=.BTF .btf.vmlinux.bin .btf.vmlinux.bin.
>
>Can you try to see where your x86_64 is coming from?

llvm-objdump -f does not print bfdarch (ARCH= in binutils-gdb/ld/emulparams/*.sh).

% objdump -f .btf.vmlinux.bin.o

.btf.vmlinux.bin.o:     file format elf64-x86-64
architecture: i386:x86-64, flags 0x00000010:
HAS_SYMS
start address 0x0000000000000000

% llvm-objdump -f .btf.vmlinux.bin.o

.btf.vmlinux.bin.o:     file format elf64-x86-64

architecture: x86_64
start address: 0x0000000000000000

% objcopy -I binary -O elf64-x86-64 -B i386:x86-64 --rename-section .data=.BTF .btf.vmlinux.bin meow.btf.vmlinux.bin.o
# happy
% objcopy -I binary -O elf64-x86-64 -B x86-64 --rename-section .data=.BTF .btf.vmlinux.bin meow.btf.vmlinux.bin.o
objcopy: architecture x86-64 unknown


As a non-x86 example, elf64-powerpcle / powerpc:common64:

% powerpc64le-linux-gnu-objdump -f meow.btf.vmlinux.bin.o

meow.btf.vmlinux.bin.o:     file format elf64-powerpcle
architecture: powerpc:common64, flags 0x00000010:


Unfortunately, GNU objcopy<2.34 (before I complained about the redundant -B https://sourceware.org/bugzilla/show_bug.cgi?id=24968)
could not infer -B from -O elf* .
% objcopy -I binary -O elf64-x86-64 --rename-section .data=.BTF .btf.vmlinux.bin .btf.vmlinux.bin.o #<2.34
% file .btf.vmlinux.bin.o
.btf.vmlinux.bin.o: ELF 64-bit LSB relocatable, no machine, version 1 (SYSV), not stripped
objcopy: architecture x86-64 unknown

GNU ld and lld will error for e_machine==0.



I will be a bit nervous to make llvm-objdump behave more BFD like.
Adding i386:x86-64, powerpc:common64, etc does not look particularly clean.
Fortunately, looking at the code, it seems that we only want to retain .BTF
The following scheme may be simpler:

objcopy --only-section=.BTF .tmp_vmlinux.btf .btf.vmlinux.bin.o && printf '\1' | dd of=.btf.vmlinux.bin.o conv=notrunc bs=1 seek=16

The command after && is to change e_type from ET_EXEC to ET_REL. GNU ld has an extremely rare feature that allows ET_EXEC to be linked,
but lld is more rigid and will reject such an input file.
https://mail.coreboot.org/hyperkitty/list/seabios@seabios.org/thread/HHIUPUXRIZ3KLTK4TPLG2V4PFP32HRBE/

Powered by blists - more mailing lists