Re: [PATCH 1/2] Add libfuzzer target to fuzz/x86_instruction_emulator

[Jan Beulich <jbeulich@xxxxxxxx>]

On 25.06.2024 14:40, Tamas K Lengyel wrote:
> On Tue, Jun 25, 2024 at 7:52 AM Jan Beulich <jbeulich@xxxxxxxx> wrote:
>>
>> On 25.06.2024 13:12, Tamas K Lengyel wrote:
>>> On Tue, Jun 25, 2024 at 2:00 AM Jan Beulich <jbeulich@xxxxxxxx> wrote:
>>>>
>>>> On 24.06.2024 23:23, Tamas K Lengyel wrote:
>>>>> On Mon, Jun 24, 2024 at 11:55 AM Jan Beulich <jbeulich@xxxxxxxx> wrote:
>>>>>>
>>>>>> On 21.06.2024 21:14, Tamas K Lengyel wrote:
>>>>>>> @@ -58,6 +58,9 @@ afl-harness: afl-harness.o $(OBJS) cpuid.o wrappers.o
>>>>>>>  afl-harness-cov: afl-harness-cov.o $(patsubst %.o,%-cov.o,$(OBJS)) 
>>>>>>> cpuid.o wrappers.o
>>>>>>>       $(CC) $(CFLAGS) $(GCOV_FLAGS) $(addprefix 
>>>>>>> -Wl$(comma)--wrap=,$(WRAPPED)) $^ -o $@
>>>>>>>
>>>>>>> +libfuzzer-harness: $(OBJS) cpuid.o
>>>>>>> +     $(CC) $(CFLAGS) $(LIB_FUZZING_ENGINE) -fsanitize=fuzzer $^ -o $@
>>>>>>
>>>>>> What is LIB_FUZZING_ENGINE? I don't think we have any use of that in the
>>>>>> tree anywhere.
>>>>>
>>>>> It's used by oss-fuzz, otherwise it's not doing anything.
>>>>>
>>>>>>
>>>>>> I'm further surprised you get away here without wrappers.o.
>>>>>
>>>>> Wrappers.o was actually breaking the build for oss-fuzz at the linking
>>>>> stage. It works just fine without it.
>>>>
>>>> I'm worried here, to be honest. The wrappers serve a pretty important
>>>> role, and I'm having a hard time seeing why they shouldn't be needed
>>>> here when they're needed both for the test and afl harnesses. Could
>>>> you add some more detail on the build issues you encountered?
>>>
>>> With wrappers.o included doing the build in the oss-fuzz docker
>>> (ubuntu 20.04 base) fails with:
>>>
>>> ...
>>> clang -O1 -fno-omit-frame-pointer -gline-tables-only
>>> -Wno-error=enum-constexpr-conversion
>>> -Wno-error=incompatible-function-pointer-types
>>> -Wno-error=int-conversion -Wno-error=deprecated-declarations
>>> -Wno-error=implicit-function-declaration -Wno-error=implicit-int
>>> -DFUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION -fsanitize=address
>>> -fsanitize-address-use-after-scope -fsanitize=fuzzer-no-link -m64
>>> -DBUILD_ID -fno-strict-aliasing -std=gnu99 -Wall -Wstrict-prototypes
>>> -Wno-unused-but-set-variable -Wno-unused-local-typedefs   -g3 -Werror
>>> -Og -fno-omit-frame-pointer
>>> -D__XEN_INTERFACE_VERSION__=__XEN_LATEST_INTERFACE_VERSION__ -MMD -MP
>>> -MF .libfuzzer-harness.d -D_LARGEFILE_SOURCE -D_LARGEFILE64_SOURCE
>>> -I/src/xen/tools/fuzz/x86_instruction_emulator/../../../tools/include
>>> -D__XEN_TOOLS__ -iquote . -fsanitize=fuzzer -fsanitize=fuzzer
>>> -Wl,--wrap=fwrite -Wl,--wrap=memcmp -Wl,--wrap=memcpy
>>> -Wl,--wrap=memset -Wl,--wrap=printf -Wl,--wrap=putchar -Wl,--wrap=puts
>>> -Wl,--wrap=snprintf -Wl,--wrap=strstr -Wl,--wrap=vprintf
>>> -Wl,--wrap=vsnprintf fuzz-emul.o x86-emulate.o x86_emulate/0f01.o
>>> x86_emulate/0fae.o x86_emulate/0fc7.o x86_emulate/decode.o
>>> x86_emulate/fpu.o cpuid.o wrappers.o -o libfuzzer-harness
>>> /usr/bin/ld: /usr/bin/ld: DWARF error: invalid or unhandled FORM value: 0x25
>>> /usr/local/lib/clang/18/lib/x86_64-unknown-linux-gnu/libclang_rt.fuzzer.a(fuzzer.o):
>>> in function `std::__Fuzzer::__libcpp_snprintf_l(char*, unsigned long,
>>> __locale_struct*, char const*, ...)':
>>> cxa_noexception.cpp:(.text._ZNSt8__Fuzzer19__libcpp_snprintf_lEPcmP15__locale_structPKcz[_ZNSt8__Fuzzer19__libcpp_snprintf_lEPcmP15__locale_structPKcz]+0x9a):
>>> undefined reference to `__wrap_vsnprintf'
>>> clang: error: linker command failed with exit code 1 (use -v to see 
>>> invocation)
>>> make: *** [Makefile:62: libfuzzer-harness] Error 1
>>> rm x86-emulate.c wrappers.c cpuid.c
>>> make: Leaving directory '/src/xen/tools/fuzz/x86_instruction_emulator'
>>> ERROR:__main__:Building fuzzers failed.
>>
>> Hmm, yes, means we'll need an actual vsnprintf() wrapper, not just a
>> declaration thereof.
> 
> I don't really get what this wrapper accomplishes

They guard against clobbering of in-register state (SIMD registers in
particular, but going forward maybe also eGRP-s as introduced by APX)
by library functions called between emulation of individual insns (or,
especially possible for fuzzing instrumented code, I think) even from
in the middle of emulating an insn. (Something as simple as the
compiler inserting a call to memcpy() or memset() somewhere in the
translation of the emulator source code could also clobber state.)

> and as I said, fuzzing works with oss-fuzz just fine without it.

I'm inclined to take this as "it appears to work just fine". Fuzzed
input register state may be lost by doing a library call somewhere,
rendering the fuzzing results less useful. This would pretty
certainly stop being tolerable the moment you compared results of
native execution of a sequence of instructions with the emulated
counterpart.

Jan