Re: [PATCH v5] perf machine: arm/arm64: Improve completeness for kernel address space

[Adrian Hunter]

On 2/09/19 5:15 PM, Leo Yan wrote:
> Hi Adrian,
> 
> On Mon, Aug 26, 2019 at 08:51:05PM +0800, Leo Yan wrote:
>> Hi Adrian,
>>
>> On Fri, Aug 16, 2019 at 04:00:02PM +0300, Adrian Hunter wrote:
>>> On 16/08/19 4:45 AM, Leo Yan wrote:
>>>> Hi Adrian,
>>>>
>>>> On Thu, Aug 15, 2019 at 02:45:57PM +0300, Adrian Hunter wrote:
>>>>
>>>> [...]
>>>>
>>>>>>> How come you cannot use kallsyms to get the information?
>>>>>>
>>>>>> Thanks for pointing out this.  Sorry I skipped your comment "I don't
>>>>>> know how you intend to calculate ARM_PRE_START_SIZE" when you reviewed
>>>>>> the patch v3, I should use that chance to elaborate the detailed idea
>>>>>> and so can get more feedback/guidance before procceed.
>>>>>>
>>>>>> Actually, I have considered to use kallsyms when worked on the previous
>>>>>> patch set.
>>>>>>
>>>>>> As mentioned in patch set v4's cover letter, I tried to implement
>>>>>> machine__create_extra_kernel_maps() for arm/arm64, the purpose is to
>>>>>> parse kallsyms so can find more kernel maps and thus also can fixup
>>>>>> the kernel start address.  But I found the 'perf script' tool directly
>>>>>> calls machine__get_kernel_start() instead of running into the flow for
>>>>>> machine__create_extra_kernel_maps();
>>>>>
>>>>> Doesn't it just need to loop through each kernel map to find the lowest
>>>>> start address?
>>>>
>>>> Based on your suggestion, I worked out below change and verified it
>>>> can work well on arm64 for fixing up start address; please let me know
>>>> if the change works for you?
>>>
>>> How does that work if take a perf.data file to a machine with a different
>>> architecture?
>>
>> Sorry I delayed so long to respond to your question; I didn't have
>> confidence to give out very reasonale answer and this is the main reason
>> for delaying.
> 
> Could you take chance to review my below replying?  I'd like to get
> your confirmation before I send out offical patch.

It is not necessary to do kallsyms__parse for x86_64, so it would be better
to check the arch before calling that.

However in general, having to copy and use kallsyms with perf.data if on a
different arch does not seem very user friendly.  But really that is up to
Arnaldo.

> 
> Thanks,
> Leo Yan
> 
>>
>> For your question for taking a perf.data file to a machine with a
>> different architecture, we can firstly use command 'perf buildid-list'
>> to print out the buildid for kallsyms, based on the dumped buildid we
>> can find out the location for the saved kallsyms file; then we can use
>> option '--kallsyms' to specify the offline kallsyms file and use the
>> offline kallsyms to fixup kernel start address.  The detailed commands
>> are listed as below:
>>
>> root@debian:~# perf buildid-list
>> 7b36dfca8317ef74974ebd7ee5ec0a8b35c97640 [kernel.kallsyms]
>> 56b84aa88a1bcfe222a97a53698b92723a3977ca /usr/lib/systemd/systemd
>> 0956b952e9cd673d48ff2cfeb1a9dbd0c853e686 /usr/lib/aarch64-linux-gnu/libm-2.28.so
>> [...]
>>
>> root@debian:~# perf script --kallsyms ~/.debug/\[kernel.kallsyms\]/7b36dfca8317ef74974ebd7ee5ec0a8b35c97640/kallsyms
>>
>> The amended patch is as below, please review and always welcome
>> any suggestions or comments!
>>
>> diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
>> index 5734460fc89e..593f05cc453f 100644
>> --- a/tools/perf/util/machine.c
>> +++ b/tools/perf/util/machine.c
>> @@ -2672,9 +2672,26 @@ int machine__nr_cpus_avail(struct machine *machine)
>>  	return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
>>  }
>>  
>> +static int machine__fixup_kernel_start(void *arg,
>> +				       const char *name __maybe_unused,
>> +				       char type,
>> +				       u64 start)
>> +{
>> +	struct machine *machine = arg;
>> +
>> +	type = toupper(type);
>> +
>> +	/* Fixup for text, weak, data and bss sections. */
>> +	if (type == 'T' || type == 'W' || type == 'D' || type == 'B')
>> +		machine->kernel_start = min(machine->kernel_start, start);
>> +
>> +	return 0;
>> +}
>> +
>>  int machine__get_kernel_start(struct machine *machine)
>>  {
>>  	struct map *map = machine__kernel_map(machine);
>> +	char filename[PATH_MAX];
>>  	int err = 0;
>>  
>>  	/*
>> @@ -2696,6 +2713,22 @@ int machine__get_kernel_start(struct machine *machine)
>>  		if (!err && !machine__is(machine, "x86_64"))
>>  			machine->kernel_start = map->start;
>>  	}
>> +
>> +	if (symbol_conf.kallsyms_name != NULL) {
>> +		strncpy(filename, symbol_conf.kallsyms_name, PATH_MAX);
>> +	} else {
>> +		machine__get_kallsyms_filename(machine, filename, PATH_MAX);
>> +
>> +		if (symbol__restricted_filename(filename, "/proc/kallsyms"))
>> +			goto out;
>> +	}
>> +
>> +	if (kallsyms__parse(filename, machine, machine__fixup_kernel_start))
>> +		pr_warning("Fail to fixup kernel start address. skipping...\n");
>> +
>> +out:
>>  	return err;
>>  }
>>  
>>
>> Thanks,
>> Leo Yan
>