lkml.org 
[lkml]   [2017]   [Nov]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] perf parse events: Fix invalid precise_ip handling
From
Date
On 2017/11/20 15:33, Jiri Olsa wrote:
> On Wed, Nov 15, 2017 at 09:00:03AM +0800, zhangmengting wrote:
>> Hi Jiri, thanks for your detailed review, please see my comments inline.
>>
>>
>> On 2017/11/10 18:39, Jiri Olsa wrote:
>>> On Fri, Nov 10, 2017 at 04:28:37PM +0800, Mengting Zhang wrote:
>>>
>>> SNIP
>>>
>>>> diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
>>>> index 39b1596..25225f4 100644
>>>> --- a/tools/perf/util/parse-events.c
>>>> +++ b/tools/perf/util/parse-events.c
>>>> @@ -1369,6 +1369,32 @@ struct event_modifier {
>>>> int pinned;
>>>> };
>>>> +static int perf_get_max_precise_ip(void)
>>>> +{
>>>> + int max_precise_ip = 0;
>>>> + struct perf_event_attr attr = {
>>>> + .type = PERF_TYPE_HARDWARE,
>>>> + .config = PERF_COUNT_HW_CPU_CYCLES,
>>>> + };
>>>> +
>>>> + event_attr_init(&attr);
>>>> +
>>>> + attr.precise_ip = 3;
>>>> + attr.sample_period = 1;
>>>> +
>>>> + while (attr.precise_ip != 0) {
>>>> + int fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
>>>> + if (fd != -1){
>>>> + close(fd);
>>>> + break;
>>>> + }
>>>> + --attr.precise_ip;
>>>> + }
>>>> + max_precise_ip = attr.precise_ip;
>>>> +
>>>> + return max_precise_ip;
>>>> +}
>>> we already have a function for that, please check perf_event_attr__set_max_precise_ip
>> Yeah, I've checked that function. But perf_event_attr__set_max_precise_ip()
>> will change attr.precise_ip
>> into the max precise ip available.
>>
>> In this case, perf should only check whether the user-specified precise_ip
>> is greater than the max
>> precise_ip without changing it into maximum. Here, introduce
>> perf_get_max_precise_ip() to return
>> the max precise ip and do not change attr.precise_ip.
>>
>> But you reminds me that perf_get_max_precise_ip() can be simplied.
> well both do the same.. probe kernel for max precise level,
> so we can keep just one function for that

OKay, I will just keep that function for probing max precise level.

>>> also I think the precise level is not generic for all the events,
>>> so you should check it for specific perf_event_attr later, when
>>> the attr is ready, not in modifier parsing
>> You are right, and I would check it for specific perf_event_attr.
>>
>> BTW, I have a question. If the user-specified precise_ip is greater than the
>> max precise_ip, I wonder
>> whether it is better to adjust the user-specified precise_ip to the maximum
>> available.
> no, I think that user defined precise level should stay the
> way the user wants it.. we don't want more angry users ;-)

Humm, I am sorry for being unclear.
If the user defined precise level is greater than the max precise level,
I think there are two ways to deal with it.
1. return EINVAL to indicate the invalid precise_ip setting;
2. adjust to the max precise level available and give message to
indicate the adjustment.

Since we should check user-defined precise level in perf_evsel__config(),
when the attr is ready, I think there is a problem with method 1, if we
keep the
user defined precise level stay the way the user wants it.

With method 1, we have to let perf_evsel__config() return value and show
errno.
And this change will affect many related functions, such as
perf_evlist__config(), and files.

With method 2, we don't need to change the return type of
perf_evsel__config().

Am I right?
>
> jirka
>
> .
>


\
 
 \ /
  Last update: 2017-11-21 09:30    [W:0.053 / U:7.292 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site