Re: [PATCH v13 2/8] perf evlist: Introduce aux evlist

["Wangnan (F)" <wangnan0@huawei.com>]

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.