lkml.org 
[lkml]   [2016]   [Jul]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v13 2/8] perf evlist: Introduce aux evlist
From
Date


On 2016/7/8 22:46, Jiri Olsa wrote:
> On Wed, Jul 06, 2016 at 08:16:52PM +0800, Wangnan (F) wrote:
>>
>> On 2016/7/6 19:36, Jiri Olsa wrote:
>>> On Mon, Jul 04, 2016 at 06:20:03AM +0000, Wang Nan wrote:
>>>
>>> SNIP
>>>
>>>> +struct perf_evlist *perf_evlist__new_aux(struct perf_evlist *parent)
>>>> +{
>>>> + struct perf_evlist *evlist;
>>>> +
>>>> + if (perf_evlist__is_aux(parent)) {
>>>> + pr_err("Internal error: create aux evlist from another aux evlist\n");
>>>> + return NULL;
>>>> + }
>>>> +
>>>> + evlist = zalloc(sizeof(*evlist));
>>>> + if (!evlist)
>>>> + return NULL;
>>>> +
>>>> + perf_evlist__init(evlist, parent->cpus, parent->threads);
>>>> + evlist->parent = parent;
>>>> + INIT_LIST_HEAD(&evlist->list);
>>>> + list_add(&evlist->list, &parent->children);
>>> I understand there's some reason for separating maps with and
>>> without overwrite set, but I'm missing it.. why is that?
>> You are asking overwrite, not write_backward?
>>
>> Overwrite mapping needs to be mapped without PROT_WRITE, so its
>> control page is also read only, so perf_evlist__mmap_consume() is
>> not able to use, and there's no way to tell kernel to where we have
>> read. Kernel overwrite old records when its full. Compare with normal
>> mapping: perf uses perf_evlist__mmap_consume() to tell kernel the
>> last byte it has read, so kernel stop writing data to it when it full,
>> and issues LOST event. This is the reason we need to separate maps
>> with and without overwrite set.
>>
>> For write backward: kernel write data in different direction, so
>> requires map separation.
> I dont like the idea of duplicating whole perf_evlist
> in order just to map some events with overwrite/backward
>
> perf_evlist carries all the other info about events,
> not just memory maping..
>
> I think it'd be better to do it some other way, like:
>
> - we have mmaps for events/evsels, so you're able to map
> it differently with or without PROT_WRITE even in current
> design.. there's struct perf_mmap that can carry that info
> then it's the matter of reading/processing those maps
> that needs to change.. new perf_evlist interface
>
> - we could keep separate struct perf_mmap arrays for forward
> and backward/overwrite maps
>
> - ...
>
> I understand both mapping need different treatment,
> but I think that should be encapsulated within the
> struct perf_evlist interface

I don't like it either, but aux_evlist is the easiest way to
do this work. Other potential solutions require heavy API changes.

Thank you.

\
 
 \ /
  Last update: 2016-07-11 13:01    [W:5.569 / U:0.028 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site