Message-ID: <4CDD22EF.4040007@hitachi.com>
Date: Fri, 12 Nov 2010 20:20:15 +0900
From: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Organization: Systems Development Lab., Hitachi, Ltd., Japan
User-Agent: Mozilla/5.0 (Windows; U; Windows NT 5.1; ja; rv:1.9.2.12) Gecko/20101027 Thunderbird/3.1.6
MIME-Version: 1.0
To: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Thomas Gleixner <tglx@linutronix.de>, Steven Rostedt <rostedt@goodmis.org>,
        Peter Zijlstra <peterz@infradead.org>,
        Frederic Weisbecker <fweisbec@gmail.com>,
        Arnaldo Carvalho de Melo <acme@infradead.org>,
        David Sharp <dhsharp@google.com>,
        Arjan van de Ven <arjan@infradead.org>,
        Andrew Morton <akpm@linux-foundation.org>, Ingo Molnar <mingo@elte.hu>,
        linux-kernel@vger.kernel.org, 2nddept-manager@sdl.hitachi.co.jp
Subject: Re: [RFD tracing] Tracing ABI Work Plan
References: <20101111004612.GB32564@Krystal> <4CDBA251.4010505@hitachi.com> <20101111130212.GD3992@Krystal>
In-Reply-To: <20101111130212.GD3992@Krystal>
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 8329
Lines: 181

(2010/11/11 22:02), Mathieu Desnoyers wrote:
> * Masami Hiramatsu (masami.hiramatsu.pt@hitachi.com) wrote:
>> Hi,
>>
>> (2010/11/11 9:46), Mathieu Desnoyers wrote:
>>> A) New ABI for user-space
>>>
>>> This new ABI will provide the features long-awaited by tracing users. We've
>>> already started this discussion by listing the requirements and acknowledging
>>> them. It is now time to start discussing the ABI. Upon disagreement on answering
>>> a specific requirement, two questions will arise:
>>>
>>> 1. How much trouble it really is to take care of this requirement. If the answer
>>>    is "not much", then we simply take care of it.
>>> 2. If it really is a big deal to take care of a requirement at the ABI level,
>>>    then we will have to discuss the use-cases.
>>>
>>> Once we are on the same page with respect to these requirements, we can come up
>>> with an ABI proposal for:
>>>
>>> - Tracing control
>>> - Trace format
>>>
>>>
>>> B) Internal Instrumentation API
>>>
>>> I propose to standardize the instrumentation mechanisms (Tracepoints, Function
>>> Tracer, Dynamic Probes, System Call Tracing, etc), so they can be used by
>>> Ftrace, Perf, and by the new ABI without requiring to build all three tracer
>>> ABI code-bases in the kernel. This involves modularizing the instrumentation
>>> sources, taking the following aspects into account:
>>>
>>> - They should be stand-alone objects, which can be built without a full tracer
>>>   enabled.
>>> - They should offer a "registration/unregistration" API to tracers, so tracers
>>>   can register a callback along with a private data pointer (this would fit
>>>   with the "multiple concurrent tracing session" requirement).
>>> - They should call these callbacks passing the private data pointer when the
>>>   instrumentation is hit.
>>> - They should provide a mechanism to list the available instrumentation (when it
>>>   makes sense) and active instrumentation. E.g., it makes sense for tracepoints
>>>   to list the available tracepoints, but it only makes sense for dynamic probes
>>>   to list the enabled probes.
>>>
>>> Masami Hiramatsu and Frederic Weisbecker already showed interest in undertaking
>>> this task.
>>
>> Actually, I didn't talked about what API should be provided internally.
>> (Yeah, I know LTTng handler want that. However, there is no "external" handler
>>  _inside_ linux kernel tree)
> 
> My target here is not LTTng. My goal is to get the ball rolling for the improved
> ABI. If we make sure all instrumentation sources provide a clean API to Ftrace,
> Perf, and eventually the new ABI, then it makes it easier to transition from one
> ABI to another; we would not have to change the "whole world", but rather just
> to switch to the new ABI when it is deemed ready.

Hmm, would you mean that staying perf and ftrace internal code separated and
adding new tracing ABI?? If so, I don't like that. Even now we've already had
two handlers for one event, adding one more is a nightmare! :(
Instead, I'd like to add PMU/HWBP/MMIO etc. to ftrace as trace events,
this will be done step by step. (Those might have two handlers)
If you or Steven or Peter integrate the ring-buffer ABI, those events may
just move onto the new ABI.
Of course, each instrumentation will have "enable/disable" method for
fitting to trace-event interface, and it will be different implementation.

>> Instead, I and Frederic talked shortly about something like user interface
>> for events. (so it would be more close to A, about controlling)
> 
> Yep, this too makes sense.
> 
>> As Thomas said, eventually kernel internal tracer should simply provide
>> "events tracing" functionality. User tools will analyze it and it's not
>> kernel's business. I agree with his opinion.
> 
> Right.
> 
>> From above viewpoint, currently only trace-events(tracepoint-based events)
>> and dynamic-events (kprobe-based events) are providing same interface for
>> users. And, for example, perf's PMU events or ftrace's mcount events aren't
>> shown up under debugfs/tracing/events. IMHO, all events provided by kernel
>> should be there, so that user tools can read the format and control those
>> events same way.
> 
> We should decide if we keep this stuff under /debugfs or move it elsewhere. This
> is part of the ABI after all.

Indeed, however at kernel summit, we decided put stable events under
/sys/kernel/events and unstables under debugfs/events, didn't we?


> Independently of where this ends up, the
> operations we need to perform are:
> 
> - For each instrumentation source (tracepoints, function tracing, dynamic
>   probes, PMC, ...)
>   - List available instrumentation
>     - Makes sense for tracepoints and PMC, but function graph tracer and dynamic
>       probes might skip this part.

I think debugfs is enough flexible;
$ for ev in `ls -d /debugfs/events/*/*`; do [ -d $ev ] && basename $ev; done

>   - List activated instrumentation

Ditto.
$ for ev in `ls -d /debugfs/events/*/*`; do [ -f $ev/enable ] && \
 [ `cat $ev/enable` -ne 0 ] && basename $ev; done

>   - Control interface
>     - Activate/deactivate instrumentation, on a per trace session basis
>       - Note: the concept of "trace session" is currently inexisting in both
>         perf and ftrace. We'll have to think of something in terms of ABI here.

Right, until ring buffer is integrated, it just use ftrace buffer.
After integrated, if each ring buffer has unique rb-id, we can use it
for "enable" event;

$ echo 4 >> /debugfs/events/foo/bar/enable
$ cat /debugfs/events/foo/bar/enable
1
3
4

But, anyway this will be done with integrating ring buffer. Currently,
this "enable" knob is only for ftrace.

>       - Note2: each instrumentation source will expects its own sets of
>         parameters to specify the instrumentation to control.

For this purpose, I'd like to introduce "virtual events".
User defines new virtual events with parameters, e.g. watching address
for harware breakpoint, function-filter for function-tracer. It will
work like a bookmark of event. Those parameters will be set when
the event is enabled (activated). Of course, some events may conflict
each other, in that case, enabling the event just returns -EBUSY.

>       - Note3: Handling of instrumentation in dynamically loadable modules
>         (which applies also to dynamic probes) might require that we allow the
>         control interface to activate a tracepoint or dynamic probe for a trace
>         session (e.g. by name) before the instrumentation point is listed as
>         available instrumentation. The goal is to deal with modules dynamically
>         loaded and dynamic instrumentation dynamically added while the trace is
>         being recorded; without requiring any user knowledge about
>         module-specific parameters whatsoever.

This should be done, but not needed for the first step.

>> For this purpose, I'd like to expand trace-event/dynamic-event framework to
>> those events. It seems that some PMU events can be treated as trace-events,
>> mcount and other parametric events can be treated as dynamic-events.
>>
>> Anyway, those stuffs can be done without new-ring-buffer-ABI things.
>> I'll just expand dyn-events a bit far from here :-)
> 
> Steven wanted to clean up his debugfs event description files, so this would fit
> well with this effort, and is indeed an ABI change. One way to do it is to keep
> the old files around and come up with a new hierarchy for the "cleaned up"
> files, along with the new features you want to add.
> 
> Also, we might want to consider moving the debugfs event description files to a
> slightly different format (see my metadata proposal). It expands a bit on the
> current information, and allows us to deal with bitfields much more elegantly.
> However this is also an ABI change.

Hmm, I think format change is another issue. Currently, no one needs different
format. We can go forward step by step. :)

Thank you,

-- 
Masami HIRAMATSU
2nd Dept. Linux Technology Center
Hitachi, Ltd., Systems Development Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/