Message-ID: <90265D59-C5B9-4AD2-B5EA-0ADC9BEF7C79@fb.com>
Date:   Mon, 11 Feb 2019 20:30:22 +0000
From:   Song Liu <songliubraving@...com>
To:     Stephane Eranian <eranian@...gle.com>
CC:     Arnaldo Carvalho de Melo <acme@...nel.org>,
        Jiri Olsa <jolsa@...hat.com>,
        Alexey Budankov <alexey.budankov@...ux.intel.com>,
        Jiri Olsa <jolsa@...nel.org>,
        lkml <linux-kernel@...r.kernel.org>,
        Ingo Molnar <mingo@...nel.org>,
        Namhyung Kim <namhyung@...nel.org>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Peter Zijlstra <a.p.zijlstra@...llo.nl>,
        Adrian Hunter <adrian.hunter@...el.com>,
        Andi Kleen <ak@...ux.intel.com>
Subject: Re: [RFC/PATCH 00/14] perf record: Add support to store data in
 directory



> On Feb 11, 2019, at 11:30 AM, Stephane Eranian <eranian@...gle.com> wrote:
> 
> Arnaldo,
> 
> On Mon, Feb 11, 2019 at 10:55 AM Arnaldo Carvalho de Melo
> <acme@...nel.org> wrote:
>> 
>> Em Mon, Feb 11, 2019 at 10:34:16AM -0800, Stephane Eranian escreveu:
>>> Jiri,
>>> 
>>> On Mon, Feb 11, 2019 at 2:20 AM Jiri Olsa <jolsa@...hat.com> wrote:
>>>> 
>>>> On Tue, Feb 05, 2019 at 02:37:27PM +0100, Jiri Olsa wrote:
>>>>> On Mon, Feb 04, 2019 at 02:44:37PM -0800, Stephane Eranian wrote:
>>>>>> Jiri,
>>>>>> 
>>>>>> While you're looking at the output format, I think it would be good
>>>>>> time to simplify the code handling perf.data file.
>>>>>> Today, perf record can emit in two formats: file mode or pipe mode.
>>>>>> This adds complexity in the code and
>>>>>> is error prone as the file mode path is tested more than the pipe mode
>>>>>> path. We have run into multiple issues with
>>>>>> the pipe mode in recent years. There is no real reason why we need to
>>>>>> maintain two formats. If I recall, the pipe format
>>>>>> was introduced because on pipes you cannot lseek to update the headers
>>>>>> and therefore some of the information present as tables
>>>>>> updated on the fly needed to be generated as pseudo records by the
>>>>>> tool. I believe that the pipe format covers all the needs and could
>>>>>> supersede the file mode format. That would simplify code in perf
>>>>>> record and eliminate the risk of errors when new headers
>>>>>> are introduced.
>>>>> 
>>>>> yep, I think we have almost all the features covered for pipe mode,
>>>>> and we have all necessary events to describe events features
>>>>> 
>>>>> so with some effort we could switch off the superfluos file header
>>>>> and use only events to describe events ;-) make sense, I'll check
>>>>> on it
>>>> 
>>>> so following features are not synthesized:
>>>> 
>>>>        FEAT_OPN(TRACING_DATA,  tracing_data,   false),
>>>>        FEAT_OPN(BUILD_ID,      build_id,       false),
>>>>        FEAT_OPN(BRANCH_STACK,  branch_stack,   false),
>>>>        FEAT_OPN(AUXTRACE,      auxtrace,       false),
>>>>        FEAT_OPN(STAT,          stat,           false),
>>>>        FEAT_OPN(CACHE,         cache,          true),
>>>> 
>>> What do you need for BRANCH_STACK?
>>> 
>>>> I think all could be added and worked around with exception
>>>> of BUILD_ID, which we store at the end (after processing
>>>> all data) and we need it early in the report phase
>>>> 
>>> Buildids are injected after the fact via perf inject when in pipe mode.
>>> 
>>>> maybe it's time to re-think that buildid -> mmap event
>>>> association again, because it's pain in current implementation
>>>> as well
>>>> 
>>> Sure, but what do you propose?
>> 
>> this keeps resurfacing, the idea is to have the building go together
>> with the PERF_RECORD_MMAP3 event, i.e. as part of setting up an
>> executable mapping the loader would get the buildid and ask the kernel
>> to keep it aroung, then when a PERF_RECORD_MMAP needs to be issued, it
>> can include the build id, so tooling will not need to get it.
>> 
> And how would the dynamic loader (ld.so) communicate the buildid to the kernel?
> How would that work for statically linked binaries.
> I think you're say the kernel would parse the ELF header looking for
> that note section
> and extract the buildid from there. Is that what you are proposing?

We have kernel parses ELF header for BUILD-ID in BPF side. You can 
find the code in stack_map_get_build_id_offset() and functions called
by it. 

> 
>> Alternatively, we would have a separate thread to process
>> PERF_RECORD_MMAP events, and as soon as it gets one from the kernel,
>> augment it straight away with the build-id it reads from the ELF file,
>> i.e. no need to have the kernel provide it, do it just like we do with
>> PERF_RECORD_BPF_EVENT, which reminds me Song probably already posted
>> thise bits...
>> 
> But that would not work in pipe mode, wouldn't it?
> Unless that thread intercepts everything pushed to the pipe looking
> for MMAP records.

For PERF_RECORD_BPF_EVENT, I am adding a separate thread, which only
listen to PERF_RECORD_BPF_EVENT with watermark of 1. This means, 
each PERF_RECORD_BPF_EVENT is sent to two ring buffers. One of them
got written to the pipe, the other is only processed by the listening
thread. Please see https://patchwork.ozlabs.org/patch/1039091/ for 
details. 

Thanks,
Song

> 
>>>> looks like bpf code is actualy getting build ids and storing
>>>> it for the callchains in kernel.. we can check if we can do
>>>> something similar for mmap events
>>>> 
>>>> jirka
>> 
>> --
>> 
>> - Arnaldo

Powered by blists - more mailing lists