Re: [PATCH v1 0/6] Simplify linking against tools/perf code

[Adrian Hunter]

On 28/03/23 20:42, Ian Rogers wrote:
> On Tue, Mar 28, 2023 at 10:12 AM Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>>
>> On 28/03/23 19:14, Ian Rogers wrote:
>>> On Tue, Mar 28, 2023 at 6:24 AM Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>>>>
>>>> On 28/03/23 04:40, Ian Rogers wrote:
>>>>> When fuzzing something like parse-events, having the main function in
>>>>> perf.c alongside global variables like input_name means that
>>>>> input_name must be redeclared with the fuzzer function's
>>>>> main. However, as the fuzzer is using the tools/perf code as a library
>>>>> this causes backward linking reference that the linker may warn
>>>>> about. Reorganize perf.c and perf.h to avoid potential backward
>>>>> references, or so that the declaration/definition locations are more
>>>>> consistent.
>>>>>
>>>>
>>>> Seems like it could be a pain to maintain.
>>>>
>>>> Did you consider just adding:
>>>>
>>>> diff --git a/tools/perf/perf.c b/tools/perf/perf.c
>>>> index 82bbe0ca858b..a75dd47d68ee 100644
>>>> --- a/tools/perf/perf.c
>>>> +++ b/tools/perf/perf.c
>>>> @@ -456,6 +456,7 @@ static int libperf_print(enum libperf_print_level level,
>>>>         return veprintf(level, verbose, fmt, ap);
>>>>  }
>>>>
>>>> +#ifndef CUSTOM_MAIN
>>>>  int main(int argc, const char **argv)
>>>>  {
>>>>         int err;
>>>> @@ -576,3 +577,4 @@ int main(int argc, const char **argv)
>>>>  out:
>>>>         return 1;
>>>>  }
>>>> +#endif
>>>>
>>>
>>> It's possible. Would need to make the static functions not warn about
>>> being declared and not used. I still think that just aligning
>>> definitions and declarations yields the most expected code and will
>>> lead to fewer problems in the long run.
>>
>> Making perf source dependent on an unknown derivative makes
>> things more complicated.
>>
>> If you are not going to contribute it to perf, then a
>> suggestion is along the lines of the following:
> 
> There's not an issue with contribution, most of the fuzzers are very
> simple. For example, here is the coresight fuzzer we run:
> 
> ```
> #include <fuzzer/FuzzedDataProvider.h>
> 
> extern "C" {
> #include "tools/perf/util/cs-etm-decoder/cs-etm-decoder.h"
> }
> 
> static void null_packet_dump(const char *) {}
> 
> extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
>  FuzzedDataProvider fdp(data, size);
>  int num_cpu = fdp.ConsumeIntegralInRange(1, 1024);
>  struct cs_etm_decoder_params d_params;
>  std::unique_ptr<struct cs_etm_trace_params[]>
>      t_params(new cs_etm_trace_params[num_cpu]);
> 
>  for(int i=0; i < num_cpu; ++i) {
>    t_params[i].protocol =
> fdp.ConsumeIntegralInRange(static_cast<int>(CS_ETM_PROTO_ETMV3)
> ,
> 
> static_cast<int>(CS_ETM_PROTO_PTM));
>    fdp.ConsumeData(&t_params[i].etmv4, sizeof(struct cs_etmv4_trace_params));
>  }
> 
>  d_params.packet_printer = null_packet_dump;
>  d_params.operation = CS_ETM_OPERATION_DECODE;
>  d_params.data = NULL;
>  d_params.formatted = true;
>  d_params.fsyncs = false;
>  d_params.hsyncs = false;
>  d_params.frame_aligned = true;
> 
>  std::unique_ptr<struct cs_etm_decoder, decltype(&cs_etm_decoder__free)>
>      decoder(cs_etm_decoder__new(num_cpu, &d_params, t_params.get()),
>              cs_etm_decoder__free);
> 
>  if (decoder == nullptr)
>    return 0;
> 
>  do {
>    size_t consumed;
>   uint8_t buf[1024];
>    size_t len = fdp.ConsumeData(buf, sizeof(buf));
> 
>    int ret = cs_etm_decoder__process_data_block(
>        decoder.get(), 0, buf, len, &consumed);
>    if (ret)
>      return 0;
> 
>  } while (fdp.remaining_bytes() > 0);
> 
>  return 0;
> }
> ```
> Most of the code is boiler plate around a single function call. The
> issues are build support, where to put the fuzzer generated test
> corpus, the code is C++, etc.
> 
>> diff --git a/tools/perf/perf.c b/tools/perf/perf.c
>> index 82bbe0ca858b..6a7fe1534664 100644
>> --- a/tools/perf/perf.c
>> +++ b/tools/perf/perf.c
>> @@ -456,7 +456,18 @@ static int libperf_print(enum libperf_print_level level,
>>         return veprintf(level, verbose, fmt, ap);
>>  }
>>
>> $ git diff
>> +#ifdef CUSTOM_MAIN
>> +int main(void)
>> +{
>> +       printf("This is not perf\n");
>> +       return 0;
>> +}
>> +
>> +int perf_main(int argc, const char **argv);
>> +int perf_main(int argc, const char **argv)
>> +#else
>>  int main(int argc, const char **argv)
>> +#endif
>>  {
>>         int err;
>>         const char *cmd;
>> $ make EXTRA_CFLAGS="-DCUSTOM_MAIN" NO_BPF_SKEL=1 -C tools/perf >/dev/null
>> Warning: Kernel ABI header at 'tools/include/uapi/linux/in.h' differs from latest version at 'include/uapi/linux/in.h'
>> Warning: Kernel ABI header at 'tools/arch/x86/include/asm/cpufeatures.h' differs from latest version at 'arch/x86/include/asm/cpufeatures.h'
>> Warning: Kernel ABI header at 'tools/arch/arm64/include/uapi/asm/perf_regs.h' differs from latest version at 'arch/arm64/include/uapi/asm/perf_regs.h'
>> Warning: Kernel ABI header at 'tools/include/linux/coresight-pmu.h' differs from latest version at 'include/linux/coresight-pmu.h'
>> Makefile.config:587: No sys/sdt.h found, no SDT events are defined, please install systemtap-sdt-devel or systemtap-sdt-dev
>> Makefile.config:805: Missing perl devel files. Disabling perl scripting support, please install perl-ExtUtils-Embed/libperl-dev
>> Makefile.config:1046: No libbabeltrace found, disables 'perf data' CTF format support, please install libbabeltrace-dev[el]/libbabeltrace-ctf-dev
>> Makefile.config:1075: No alternatives command found, you need to set JDIR= to point to the root of your Java directory
>> Makefile.config:1137: libpfm4 not found, disables libpfm4 support. Please install libpfm4-dev
>> $ tools/perf/perf version
>> This is not perf
>>
> 
> Sure, I'm aware of how the #ifdef would function. What I'm saying is
> that even with a CUSTOM_MAIN ifdef, which wouldn't be necessary with
> my patches as you can drop the object file, the refactoring in the
> patches still makes sense.

I am just going off what your commit messages say

> 
>   perf ui: Move window resize signal functions
> 
> It isn't clear why these UI functions should be in perf.c given window
> resizing is a UI issue.
> 
>   perf usage: Move usage strings
> 
> The usage strings were in builtin.h and not perf.h, so the
> declaration/definition didn't align. util.h declares usage(), although
> the definition is in usage.c, the variables are moved to match this.
> 
>   perf header: Move perf_version_string declaration
> 
> perf_version_string is defined in header.c and so declaring it in
> header.h rather than perf.h is consistent.
> 
>   perf version: Use regular verbose flag
> 
> Just remove an unnecessary global by repurposing the existing global
> for the job.
> 
>   perf util: Move input_name to util
> 
> There's no reference to input_name in perf.c and so util.c is as
> sensible a location. Having this be a global is something of an
> encapsulation fail, but cleaning that up wasn't the point of these
> patches. The variable is used in util, so reaching up a directory to
> get it feels somewhat unnatural.
> 
>   perf util: Move perf_guest/host declarations
> 
> As with input_name.
> 
> Thanks,
> Ian