[RFD tracing] Tracing ABI Work Plan

[RFD tracing] Tracing ABI Work Plan

[Mathieu Desnoyers]

Following the Kernel Summit 2010 and Linux Plumber Conference, we came up to
the following conclusions (we're finally making progress !!):

- The ABIs of Ftrace and Perf are in active use, thus Linus has stated
  that they can not be changed.
- It has been stated at Kernel Summit that new ABIs should not go into
  mainline first, but should be tested in an out-of-tree first, and get
  feed back from their users.
- The goal is to have both Perf and Ftrace work well together, and have
  a more generic ABI. But the current ABI of both Perf and Ftrace fail to
  meet each others requirements, as well as requirement from a large user-base.

Therefore, as Thomas Gleixner pointed out, we have two choices: either we stop
there, keep the current ABIs and give up, or we create a new set of ABIs. I
hereby propose to come up with a new set of tracing ABIs and, this time, that we
_take our time_ before mainlining them -- trying them out in a separate tree for
a while beforehand. There is no hurry to push anything to mainline anyway,
because Ftrace and Perf answer the most pressing kernel developer tracing needs.
Let's work on the new ABI at the pace that everyone is happy with it.


First of all, I'm not a big fan of "hide-and-seek" games regarding motives and
end-goals, especially when it comes to designing ABIs. So here is my entire
high-level roadmap towards having tracer fully usable by my user-base in the
mainline kernel. I propose to split the upcoming task in the following 4
sub-items.


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.


C) Ring Buffer

I propose to create a "lite" version of my generic ring buffer library which
will be specifically focused on the requirement list. It will simplify the code
review process.


D) Trace Clocks (hypervisor, kernel, user-space)

Dealing with tracing clock sources at the kernel, userland and hypervisor level
can be achieved by exposing the user requirements, deciding which approach makes
sense for synchronized trace clocks across hypervisor/kernel/userland, and
implement the kernel trace clocks, vDSO, and hypervisor support accordingly.


Feedback is welcome,

Thanks,

Mathieu

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com

Re: [RFD tracing] Tracing ABI Work Plan

[Thomas Gleixner]

Mathieu,

On Wed, 10 Nov 2010, Mathieu Desnoyers wrote:

> Following the Kernel Summit 2010 and Linux Plumber Conference, we came up to
> the following conclusions (we're finally making progress !!):
>
> - The ABIs of Ftrace and Perf are in active use, thus Linus has stated
>   that they can not be changed.
> - It has been stated at Kernel Summit that new ABIs should not go into
>   mainline first, but should be tested in an out-of-tree first, and get
>   feed back from their users.
> - The goal is to have both Perf and Ftrace work well together, and have
>   a more generic ABI. But the current ABI of both Perf and Ftrace fail to
>   meet each others requirements, as well as requirement from a large user-base.
> 
> Therefore, as Thomas Gleixner pointed out, we have two choices: either we stop
> there, keep the current ABIs and give up, or we create a new set of ABIs. I
> hereby propose to come up with a new set of tracing ABIs and, this time, that we
> _take our time_ before mainlining them -- trying them out in a separate tree for
> a while beforehand. There is no hurry to push anything to mainline anyway,
> because Ftrace and Perf answer the most pressing kernel developer tracing needs.
> Let's work on the new ABI at the pace that everyone is happy with it.

Using the observed progress rate so far that means pace is infinite
slow.

> First of all, I'm not a big fan of "hide-and-seek" games regarding motives and
> end-goals, especially when it comes to designing ABIs. So here is my entire
> high-level roadmap towards having tracer fully usable by my user-base in the

Roadmap? You failed to provide the corresponding spreadsheets, project
plans and milestones along with that.

> mainline kernel. I propose to split the upcoming task in the following 4
> sub-items.
> 
> 
> 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.

Sorry for being sarcastic, but that reads like it was adopted from the
worst possible management course handout. You forgot to mention that
the process should involve rating matrixes and other state of the art
evaluation methods.

> 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

That's going to happen somewhere around 2050, right? That's good as I
won't be in your way then anymore.

<Snipping the other proposals as I need some sleep>

I just want to tell you once more, that these marvelous writeups and
proposals which we have seen repeatedly in the last years do not solve
anything. Quite the contrary, they make folks loose interest, cause
grumpiness and prevent any useful outcome.

The time wasted on all these managerial proposals and work plans which
we have seen come by in the last years plus the resulting fruitless
discussions would have been sufficient to write at least two
generations of useful tracers.

The requirement list has been beaten to death several times already
along with the various options of trace formats, so we really are at
the point where you folks need to sit down and come up with real code
which can be discussed and improved on a technical base.

Either that or you immediately install the "Work Plan Committee" which
will make sure that there will be no progress ever.

Thanks,

	tglx

Re: [RFD tracing] Tracing ABI Work Plan

[Steven Rostedt]

On Thu, 2010-11-11 at 02:33 +0100, Thomas Gleixner wrote:

> Either that or you immediately install the "Work Plan Committee" which
> will make sure that there will be no progress ever.

Oooo ooo oo, I want to be on this committee!!!

Get Mathieu and I there, and we can guarantee no work will ever
happen ;-)

-- Steve

Re: [RFD tracing] Tracing ABI Work Plan

[Thomas Gleixner]

On Wed, 10 Nov 2010, Steven Rostedt wrote:
> On Thu, 2010-11-11 at 02:33 +0100, Thomas Gleixner wrote:
> 
> > Either that or you immediately install the "Work Plan Committee" which
> > will make sure that there will be no progress ever.
> 
> Oooo ooo oo, I want to be on this committee!!!
> 
> Get Mathieu and I there, and we can guarantee no work will ever
> happen ;-)

Hereby you and Mathieu are appointed to the "Work Plan Committee". By
accepting that appointment you promise to stay away from real kernel
tracing problems forever and restrict yourself to an annual "no
progress" report.

Thanks,

	tglx
---
P.S.: It would be more funny, if it wouldn't be close to reality!

Re: [RFD tracing] Tracing ABI Work Plan

[Steven Rostedt]

On Thu, 2010-11-11 at 02:55 +0100, Thomas Gleixner wrote:
> On Wed, 10 Nov 2010, Steven Rostedt wrote:
> > On Thu, 2010-11-11 at 02:33 +0100, Thomas Gleixner wrote:
> > 
> > > Either that or you immediately install the "Work Plan Committee" which
> > > will make sure that there will be no progress ever.
> > 
> > Oooo ooo oo, I want to be on this committee!!!
> > 
> > Get Mathieu and I there, and we can guarantee no work will ever
> > happen ;-)
> 
> Hereby you and Mathieu are appointed to the "Work Plan Committee". By
> accepting that appointment you promise to stay away from real kernel
> tracing problems forever and restrict yourself to an annual "no
> progress" report.
> 
> Thanks,
> 
> 	tglx
> ---
> P.S.: It would be more funny, if it wouldn't be close to reality!

So be it...

Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

diff --git a/MAINTAINERS b/MAINTAINERS
index 0094224..96850d6 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -5889,7 +5889,6 @@ S:	Maintained
 F:	drivers/char/tpm/
 
 TRACING
-M:	Steven Rostedt <rostedt@goodmis.org>
 M:	Frederic Weisbecker <fweisbec@gmail.com>
 M:	Ingo Molnar <mingo@redhat.com>
 T:	git git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip.git perf/core

Re: [RFD tracing] Tracing ABI Work Plan

[Thomas Gleixner]

On Wed, 10 Nov 2010, Steven Rostedt wrote:
> On Thu, 2010-11-11 at 02:55 +0100, Thomas Gleixner wrote:
> > On Wed, 10 Nov 2010, Steven Rostedt wrote:
> > > On Thu, 2010-11-11 at 02:33 +0100, Thomas Gleixner wrote:
> > > 
> > > > Either that or you immediately install the "Work Plan Committee" which
> > > > will make sure that there will be no progress ever.
> > > 
> > > Oooo ooo oo, I want to be on this committee!!!
> > > 
> > > Get Mathieu and I there, and we can guarantee no work will ever
> > > happen ;-)
> > 
> > Hereby you and Mathieu are appointed to the "Work Plan Committee". By
> > accepting that appointment you promise to stay away from real kernel
> > tracing problems forever and restrict yourself to an annual "no
> > progress" report.
> > 
> > Thanks,
> > 
> > 	tglx
> > ---
> > P.S.: It would be more funny, if it wouldn't be close to reality!
> 
> So be it...

NAK. Just running away is not a technical solution.

> Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
> 
> diff --git a/MAINTAINERS b/MAINTAINERS
> index 0094224..96850d6 100644
> --- a/MAINTAINERS
> +++ b/MAINTAINERS
> @@ -5889,7 +5889,6 @@ S:	Maintained
>  F:	drivers/char/tpm/
>  
>  TRACING
> -M:	Steven Rostedt <rostedt@goodmis.org>
>  M:	Frederic Weisbecker <fweisbec@gmail.com>
>  M:	Ingo Molnar <mingo@redhat.com>
>  T:	git git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip.git perf/core
> 
> 

[RFC tracing] Common Trace Format for Linux (v1.1)

[Mathieu Desnoyers]

* Thomas Gleixner (tglx@linutronix.de) wrote:
[...]

> The requirement list has been beaten to death several times already
> along with the various options of trace formats, so we really are at
> the point where you folks need to sit down and come up with real code
> which can be discussed and improved on a technical base.

I understand and share your frustration about things having been standing still
for way too long.

In order to get things rolling, I hereby append my trace format proposal as RFC.
I did implement the core elements of it already in the BabelTrace trace
converter project, so it's not one of these dreaded "design by committee without
any understanding of the practicality aspects" standards. I try make sure it
translates to something realistic and useful.

Feedback is welcome, thanks,

Mathieu

RFC: Common Trace Format Proposal for Linux (v1.1)

Mathieu Desnoyers, EfficiOS Inc.

The goal of the present document is to propose a trace format that suits the
needs of the embedded, telecom, high-performance and kernel communities.  It is
based on the Common Trace Format Requirements (v1.4) document. It is designed to
be natively generated by tracing of a Linux kernel and Linux user-space
applications written in C/C++.

A reference implementation of a library to read and write this trace format is
being implemented within the BabelTrace project, a converter between trace
formats. The development tree is available at:

  git tree:   git://git.efficios.com/babeltrace.git
  gitweb:     http://git.efficios.com/?p=babeltrace.git


1. Preliminary definitions

  - Trace: An ordered sequence of events.
  - Section: Group of events, containing a subset of the trace event types.
  - Packet: A sequence of physically contiguous events within a section.
  - Event: This is the basic entry in a trace. (aka: a trace record).
    - An event identifier (ID) relates to the class (a type) of event within
      a section.
        e.g. section: high_throughput, event: irq_entry.
    - An event (or event record) relates to a specific instance of an event
      class.
        e.g. section: high_throughput, event: irq_entry, at time X, on CPU Y


2. High-level representation of a trace

A trace is divided into multiple trace streams, each representing an information
stream specific to:

 - a section,
 - a processor.

A trace "section" consists of a collection of trace streams (typically one trace
stream per cpu) containing a subset of the trace event types.

Because each trace stream is appended to while a trace is being recorded, each
is associated with a separate file for disk output. Therefore, a trace stored to
disk can be represented as a directory containing one file per section.

A metadata section contains information on trace event types. It describes:

- Trace version.
- Types available.
- Per-section event header description.
- Per-section event header selection.
- Per-section event context fields.
- Per-event
  - Event type to section mapping.
  - Event type to name mapping.
  - Event type to ID mapping.
  - Event fields description.


3. Trace Section

A trace section is divided in contiguous packets of variable size. These
subdivisions allow the trace analyzer to perform a fast binary search by time
within the section (typically requiring to index only the packet headers)
without reading the whole section. These subdivisions have a variable size to
eliminate the need to transfer the packet padding when partially filled packets
must be sent when streaming a trace for live viewing/analysis. Dividing sections
into packets is also useful for network streaming over UDP and flight recorder
mode tracing (a whole packet can be swapped out of the buffer atomically for
reading).

The section header is repeated at the beginning of each packet to allow
flexibility in terms of:

  - streaming support,
  - allowing arbitrary buffers to be discarded without making the trace
    unreadable,
  - allow UDP packet loss handling by either dealing with missing packet or
    asking for re-transmission.
  - transparently support flight recorder mode,
  - transparently support crash dump.

The section header will therefore be referred to as the "packet header"
thorough the rest of this document.


4. Types

4.1 Basic types

A basic type is a scalar type, as described in this section.

4.1.1 Type inheritance

Type specifications can be inherited to allow deriving concrete types from an
abstract type. For example, see the uint32_t type derived from the "integer"
abstract type below ("Integers" section). Concrete types have a precise binary
representation in the trace. Abstract types have methods to read and write these
types, but must be derived into a concrete type to be usable in an event field.

Concrete types inherit from abstract types. Abstract types can inherit from
other abstract types.

4.1.2 Alignment

We define "byte-packed" types as aligned on the byte size, namely 8-bit.
We define "bit-packed" types as following on the next bit, as defined by the
"bitfields" section.
We define "natural alignment" of a basic type as the lesser value between the
type size and the architecture word size.

All basic types, except bitfields, are either aligned on their "natural"
alignment or byte-packed, depending on the architecture preference.
Architectures providing fast unaligned writes byte-packed basic types to save
space, aligning each type on byte boundaries (8-bit). Architectures with slow
unaligned writes align types on the lesser value between their size and the
architecture word size (the type "natural" alignment on the architecture).

Note that the natural alignment for 64-bit integers and double-precision
floating point values is fixed to 32-bit on a 32-bit architecture, but to 64-bit
for a 64-bit architecture.

Metadata attribute representation:

  align = value;                                /* value in bits */

4.1.3 Byte order

By default, target architecture endianness is used. Byte order can be overridden
for a basic type by specifying a "byte_order" attribute. Typical use-case is to
specify the network byte order (big endian: "be") to save data captured from the
network into the trace without conversion. If not specified, the byte order is
native.

Metadata representation:

  byte_order = native OR network OR be OR le;	/* network and be are aliases */

4.1.4 Size

Type size, in bits, for integers and floats is that returned by "sizeof()" in C
multiplied by CHAR_BIT.
We require the size of "char" and "unsigned char" types (CHAR_BIT) to be fixed
to 8 bits for cross-endianness compatibility.

Metadata representation:

  size = value;    (value is in bits)

4.1.5 Integers

Signed integers are represented in two-complement. Integer alignment, size,
signedness and byte ordering are defined in the metadata. Integers aligned on
byte size (8-bit) and with length multiple of byte size (8-bit) correspond to
the C99 standard integers. In addition, integers with alignment and/or size that
are _not_ a multiple of the byte size are permitted; these correspond to the C99
standard bitfields, with the added specification that the CTF integer bitfields
have a fixed binary representation. A MIT-licensed reference implementation of
the CTF portable bitfields is available at:

  http://git.efficios.com/?p=babeltrace.git;a=blob;f=include/babeltrace/bitfield.h

Binary representation of integers:

- On little and big endian:
  - Within a byte, high bits correspond to an integer high bits, and low bits
    correspond to low bits.
- On little endian:
  - Integer across multiple bytes are placed from the less significant to the
    most significant.
  - Consecutive integers are placed from lower bits to higher bits (even within
    a byte).
- On big endian:
  - Integer across multiple bytes are placed from the most significant to the
    less significant.
  - Consecutive integers are placed from higher bits to lower bits (even within
    a byte).

This binary representation is derived from the bitfield implementation in GCC
for little and big endian. However, contrary to what GCC does, integers can
cross units boundaries (no padding is required). Padding can be explicitely
added (see 4.1.6 GNU/C bitfields) to follow the GCC layout if needed.

Metadata representation:

  abstract_type integer {
    signed = true OR false;                     /* default false */
    byte_order = native OR network OR be OR le; /* default native */
    size = value;                               /* value in bits, no default */
    align = value;                              /* value in bits */
  }

Example of type inheritance (creation of a concrete type uint32_t):

type uint32_t {
  parent = integer;
  size = 8;
  signed = false;
  align = 32;
}

Definition of a 5-bit signed bitfield:

type int5_t {
  parent = integer;
  size = 5;
  signed = true;
  align = 1;
}

4.1.6 GNU/C bitfields

The GNU/C bitfields follow closely the integer representation, with a
particularity on alignment: if a bitfield cannot fit in the current unit, the
unit is padded and the bitfield starts at the following unit. We therefore need
to express the extra "unit size" information.

Metadata representation:

abstract_type gcc_bitfield {
  parent = integer;
  unit_size = value;
}

As an example, the following structure declared in C compiled by GCC:

struct example {
  short a:12;
  short b:5;
};

Would correspond to the following structure, aligned on the largest element
(short). The second bitfield would be aligned on the next unit boundary, because
it would not fit in the current unit.

type struct_example {
  parent = struct;
  fields = {
    {
      type {
        parent = gcc_bitfield;
        unit_size = 16;				/* sizeof(short) */
        size = 12;
        signed = true;
        align = 1;
      },
      a,
    },
    {
      type {
        parent = gcc_bitfield;
        unit_size = 16;				/* sizeof(short) */
        size = 5;
        signed = true;
        align = 1;
      },
      b,
    },
  };
}

4.1.7 Floating point

[ Unneeded for Linux kernel ]

4.1.8 Enumerations

Enumerations are a mapping between an integer type and a table of strings. The
numerical representation of the enumeration follows the integer type specified
by the metadata. The enumeration mapping table is detailed in the enumeration
description within the metadata.

abstract_type enum  {
  .parent = integer;
  .map = {
    { value , string },
    { value , string },
    { value , string },
    ...
  };
}


4.2 Compound types

4.2.1 Structures

Structures are aligned on the largest alignment required by basic types
contained within the structure. (This follows the ISO/C standard for structures)

Metadata representation:

abstract_type struct {
  fields = {
    { field_type, field_name },
    { field_type, field_name },
    ...
  };
}  

Example:

type struct_example {
  parent = struct;
  fields = {
    {
      type {                 /* Nameless type */
        parent = integer;
        size = 16;
        signed = true;
        align = 16;
      },
      first_field_name,
    },
    {
      uint64_t,              /* Named type declared in the metadata */
      second_field_name,
    }
  };
}

The fields are placed in a sequence next to each other. They each possess a
field name, which is a unique identifier within the structure.

4.2.2 Arrays

Arrays are fixed-length. Their length is declared in the type declaration within
the metadata. They contain an array of "inner type" elements, which can refer to
any type not containing the type of the array being declared (no circular
dependency).

Metadata representation:

abstract_type array {
  length = value;
  elem_type = type;
}

E.g.:

type example_array {
  parent = array;
  length = 10;
  elem_type = uint32_t;
}

4.2.3 Sequences

Sequences are dynamically-sized arrays. They start with an integer that specify
the length of the sequence, followed by an array of "inner type" elements.

abstract_type sequence {
  length_type = type;	/* Inheriting from integer */
  elem_type = type;
}

The integer type follows the integer types specifications, and the sequence
elements follow the "array" specifications.

4.2.4 Strings

Strings are an array of bytes of variable size and are terminated by a '\0'
"NULL" character.  Their encoding is described in the metadata. In absence of
encoding attribute information, the default encoding is UTF-8.

abstract_type string {
  encoding = UTF8 OR ASCII;
}


5. Trace Packet Header

- Aligned on page size. Fixed size. Fields aligned on their natural size or
  packed (depending on the architecture preference).
  No padding at the end of the trace packet header. Native architecture byte
  ordering.
- Magic number (CTF magic numbers: 0xC1FC1FC1 and its reverse endianness
  representation: 0xC11FFCC1) It needs to have a non-symmetric bytewise
  representation. Used to distinguish between big and little endian traces (this
  information is determined by knowing the endianness of the architecture
  reading the trace and comparing the magic number against its value and the
  reverse, 0xC11FFCC1). This magic number specifies that we use the CTF metadata
  description language described in this document. Different magic numbers
  should be used for other metadata description languages.
- Session ID, used to ensure the packet match the metadata used.
  (note: we cannot use a metadata checksum because metadata can be appended to
   while tracing is active)
- Packet content size (in bytes).
- Packet size (in bytes, includes padding).
- Packet content checksum (optional). Checksum excludes the packet header.
- Per-section packet sequence count (to deal with UDP packet loss). The number
  of significant sequence counter bits should also be present, so wrap-arounds
  are deal with correctly.
- Timestamp at the beginning and end of the packet. Should include all
  event timestamps contained therein.
- Events discarded count
  - Snapshot of a per-section free-running counter, counting the number of
    events discarded that were supposed to be written in the section prior to
    the first event in the packet.
    * Note: producer-consumer buffer full condition should fill the current
            packet with padding so we know exactly where events have been
            discarded.
- Lossless compression scheme used for the packet content. Applied directly to
  raw data.
  0: no compression scheme
  1: bzip2
  2: gzip
- Cypher used for the packet content. Applied after compression.
  0: no encryption
  1: AES
- Checksum scheme used for the packet content. Applied after encryption.
  0: no checksum
  1: md5
  2: sha1
  3: crc32

type packet_header {
  parent = struct;
  fields = {
    { uint32_t, magic },
    { uint32_t, session_id },
    { uint32_t, content_size },
    { uint32_t, packet_size },
    { uint32_t, checksum },
    { uint32_t, section_packet_count },
    { uint64_t, timestamp_begin }
    { uint64_t, timestamp_end }
    [ uint32_t, events_discarded },
    { uint8_t,  section_packet_count_bits },	/* Significant counter bits */
    { uint8_t,  compression_scheme },
    { uint8_t,  encryption_scheme },
    { uint8_t,  checksum },
  };
};


6. Event Structure

The overall structure of an event is:

  - Event Header (as specifed by the section metadata)
  - Extended Event Header (as specified by the event header)
  - Event Context (as specified by the section metadata)
  - Event Payload (as specified by the event metadata)


6.1 Event Header

One major factor can vary between sections: the number of event IDs assigned to
a section. Luckily, this information tends to stay relatively constant (modulo
event registration while trace is being recorded), so we can specify different
representations for sections containing few event IDs and sections containing
many event IDs, so we end up representing the event ID and timestamp as densely
as possible in each case.

We therefore provide two types of events headers. Type 1 accommodates sections
with less than 31 event IDs. Type 2 accommodates sections with 31 or more event
IDs.

The "extended headers" are used in the rare occasions where the information
cannot be represented in the ranges available in the event header.

Types uintX_t represent an X-bit unsigned integer.


6.1.1 Type 1 - Few event IDs

  - Aligned on 32-bit (or 8-bit if byte-packed, depending on the architecture
    preference).
  - Fixed size: 32 bits.
  - Native architecture byte ordering.

type event_header_1 {
  parent = struct;
  fields = {
    { uint5_t, id },	/*
			 * id: range: 0 - 30.
			 * id 31 is reserved to indicate a following
			 * extended header.
			 */
    { uint27_t, timestamp },
  };
};

The end of a type 1 header is aligned on a 32-bit boundary (or packed).


6.1.2 Extended Type 1 Event Header

  - Follows struct event_header_1, which is aligned on 32-bit, so no need to
    realign.
  - Fixed size: 96 bits.
  - Native architecture byte ordering.

type event_header_1_ext {
  parent = struct;
  fields = {
    { uint32_t, id },		/* 32-bit event IDs */
    { uint64_t, timestamp },	/* 64-bit timestamps */ 
  };
};

The end of a type 1 extended header is aligned on the natural alignment of a
64-bit integer (or 8-bit if byte-packed).


6.1.3 Type 2 - Many event IDs

  - Aligned on 32-bit (or 8-bit if byte-packed, depending on the architecture
    preference).
  - Fixed size: 48 bits.
  - Native architecture byte ordering.

type event_header_2 {
  parent = struct;
  fields = {
    { uint32_t, timestamp },
    { uint16_t, id },	/*
			 * id: range: 0 - 65534.
			 * id 65535 is reserved to indicate a following
			 * extended header.
			 */
  };
};

The end of a type 2 header is aligned on a 16-bit boundary (or 8-bit if
byte-packed).


6.1.4 Extended Type 2 Event Header

  - Follows struct event_header_2, which alignment end on a 16-bit boundary, so
    we need to align on 64-bit integer natural alignment (or 8-bit if
    byte-packed).
  - Fixed size: 96 bits.
  - Native architecture byte ordering.

type event_header_2_ext {
  parent = struct;
  fields = {
    { uint64_t, timestamp },	/* 64-bit timestamps */ 
    { uint32_t, id },		/* 32-bit event IDs */
  };
};

The end of a type 2 extended header is aligned on the natural alignment of a
32-bit integer (or 8-bit if byte-packed).


6.2 Event Context

The event context contains information relative to the current event. The choice
and meaning of this information is specified by the metadata "section"
information. For this trace format, event context is usually empty, except when
the metadata "section" information specifies otherwise by declaring a non-empty
structure for the event context. An example of event context is to save the
event payload size with each event, or to save the current PID with each event.

6.2.1 Event Context Description

Event context example. These are declared within the section declaration within
the metadata.

type per_section_event_ctx {
  parent = struct;
  fields = {
    { uint, pid },
    { uint16_t, payload_size },
  };
};


6.3 Event Payload

An event payload contains fields specific to a given event type. The fields
belonging to an event type are described in the event-specific metadata
within a structure type.

6.3.1 Padding

No padding at the end of the event payload. This differs from the ISO/C standard
for structures, but follows the CTF standard for structures. In a trace, even
though it makes sense to align the beginning of a structure, it really makes no
sense to add padding at the end of the structure, because structures are usually
not followed by a structure of the same type.

This trick can be done by adding a zero-length "end" field at the end of the C
structures, and by using the offset of this field rather than using sizeof()
when calculating the size of a structure (see section "A.1 Helper macros").

6.3.2 Alignment

The event payload is aligned on the largest alignment required by types
contained within the payload. (This follows the ISO/C standard for structures)



7. Metadata

The meta-data is located in a section named "metadata". It is made of "packets",
which each start with a packet header. The event type within the metadata
section have no event header nor event context. Each event only contains a
null-terminated "string" payload, which is a metadata description entry. The
events are packed one next to another. Each packet start with a packet header,
which contains, amongst other fields, the session ID and magic number.

The metadata can be parsed by reading through the metadata strings, skipping
spaces, newlines and null-characters.

trace {
  major = value;	/* Trace format version */
  minor = value;
}

section {
  name = section_name;
  event {
    /* Type 1 - Few event IDs; Type 2 - Many event IDs */
    header_type = type1 OR type2;
    context {
      event_size = true OR false;  /* Includes event size field or not */
    }
  }
}

event {
  name = event_name;
  id = value;			/* Numeric identifier within the section */
  section = section_name;
  fields = type inheriting from "struct" abstract type.
}

/* More detail on types in section 4. Types */

/* Named types */
type typename {
   ...
}

/* Unnamed types, contained within compound type fields */
type {
   ...
}

A.1 Helper macros

The two following macros keep track of the size of a GNU/C structure without
padding at the end by placing HEADER_END as the last field. A one byte end field
is used for C90 compatibility (C99 flexible arrays could be used here). Note
that this does not affect the effective structure size, which should always be
calculated with the header_sizeof() helper.

#define HEADER_END		char end_field
#define header_sizeof(type)	offsetof(typeof(type), end_field)

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com

Re: [RFC tracing] Common Trace Format for Linux (v1.1)

[Thomas Gleixner]

Mathieu,

On Wed, 10 Nov 2010, Mathieu Desnoyers wrote:

> * Thomas Gleixner (tglx@linutronix.de) wrote:
> [...]
> 
> > The requirement list has been beaten to death several times already
> > along with the various options of trace formats, so we really are at
> > the point where you folks need to sit down and come up with real code
> > which can be discussed and improved on a technical base.
> 
> I understand and share your frustration about things having been standing still
> for way too long.
> 
> In order to get things rolling, I hereby append my trace format proposal as RFC.
> I did implement the core elements of it already in the BabelTrace trace
> converter project, so it's not one of these dreaded "design by committee without
> any understanding of the practicality aspects" standards. I try make sure it
> translates to something realistic and useful.
> 
> Feedback is welcome, thanks,
> 
> Mathieu
> 
> RFC: Common Trace Format Proposal for Linux (v1.1)

Groan. Did you read what I wrote ?

> >                                              ... so we really are at
> > the point where you folks need to sit down and come up with real code
> > which can be discussed and improved on a technical base.

Your reaction on this is to send yet another proposal, which has not
really anything new in it.

What we are waiting for is a sensible incremental patch series, which
extends or replaces functionality in the existing perf ABI up to the
point, where we can eventually see the need for a sensible
replacement. That's the way we work, not with tons of proposals.

Thanks,

	tglx

Re: [RFC tracing] Common Trace Format for Linux (v1.1)

[Mathieu Desnoyers]

* Thomas Gleixner (tglx@linutronix.de) wrote:
> Mathieu,
> 
> On Wed, 10 Nov 2010, Mathieu Desnoyers wrote:
> 
> > * Thomas Gleixner (tglx@linutronix.de) wrote:
> > [...]
> > 
> > > The requirement list has been beaten to death several times already
> > > along with the various options of trace formats, so we really are at
> > > the point where you folks need to sit down and come up with real code
> > > which can be discussed and improved on a technical base.
> > 
> > I understand and share your frustration about things having been standing still
> > for way too long.
> > 
> > In order to get things rolling, I hereby append my trace format proposal as RFC.
> > I did implement the core elements of it already in the BabelTrace trace
> > converter project, so it's not one of these dreaded "design by committee without
> > any understanding of the practicality aspects" standards. I try make sure it
> > translates to something realistic and useful.
> > 
> > Feedback is welcome, thanks,
> > 
> > Mathieu
> > 
> > RFC: Common Trace Format Proposal for Linux (v1.1)
> 
> Groan. Did you read what I wrote ?
> 
> > >                                              ... so we really are at
> > > the point where you folks need to sit down and come up with real code
> > > which can be discussed and improved on a technical base.
> 
> Your reaction on this is to send yet another proposal, which has not
> really anything new in it.
> 
> What we are waiting for is a sensible incremental patch series, which
> extends or replaces functionality in the existing perf ABI up to the
> point, where we can eventually see the need for a sensible
> replacement. That's the way we work, not with tons of proposals.

OK, I'll work on this. Meanwhile, anyone interested to provide feedback on the
RFC I sent is still welcome to do so.

Thanks,

Mathieu


-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com

Re: [RFD tracing] Tracing ABI Work Plan

[Masami Hiramatsu]

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)
Instead, I and Frederic talked shortly about something like user interface
for events. (so it would be more close to A, about controlling)

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.

>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.

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 :-)

Best Regards,

-- 
Masami HIRAMATSU
2nd Dept. Linux Technology Center
Hitachi, Ltd., Systems Development Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com

Re: [RFD tracing] Tracing ABI Work Plan

[Mathieu Desnoyers]

* 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.

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

> 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.

Thanks,

Mathieu

> 
> Best Regards,
> 
> -- 
> Masami HIRAMATSU
> 2nd Dept. Linux Technology Center
> Hitachi, Ltd., Systems Development Laboratory
> E-mail: masami.hiramatsu.pt@hitachi.com

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com

Re: [RFD tracing] Tracing ABI Work Plan

[Masami Hiramatsu]

(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

Re: [RFD tracing] Tracing ABI Work Plan

[David Sharp]

On Thu, Nov 11, 2010 at 5:02 AM, Mathieu Desnoyers
<mathieu.desnoyers@efficios.com> 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.
>
>> 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. 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.
>  - List activated instrumentation
>  - 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.
>      - Note2: each instrumentation source will expects its own sets of
>        parameters to specify the instrumentation to control.
>      - 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.
>
>> 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.

Along this vein, we'd like to see a version number somewhere in the
interface. Mostly, this should version the ring buffer data headers,
event description format (not content), and control file interface
(enable, filter, etc). I think the text format that comes out of the
"trace" file doesn't necessarily need to be versioned. A simple
major.minor string would be fine.

>
> Thanks,
>
> Mathieu
>
>>
>> Best Regards,
>>
>> --
>> Masami HIRAMATSU
>> 2nd Dept. Linux Technology Center
>> Hitachi, Ltd., Systems Development Laboratory
>> E-mail: masami.hiramatsu.pt@hitachi.com
>
> --
> Mathieu Desnoyers
> Operating System Efficiency R&D Consultant
> EfficiOS Inc.
> http://www.efficios.com
>