[lttng-dev] [RELEASE] LTTng-Tools 2.11.0-rc1 - Lafontaine

Thu Sep 6 17:00:58 EDT 2018

Hi everyone!

Today marks the release of the first LTTng 2.11 - "Lafontaine" release
candidate.

This release is named after a modern Saison beer from Montréal's Oshlag
microbrewery. It is a refreshing, zesty, rice beer with hints of fruit
and spices. Some even say it makes for a great Somaek when mixed with
Chamisul Soju, not that we've tried!

Lafontaine is also a tongue-in-cheek reference to a water leak that
affected EfficiOS's offices during the development of this release.

The most notable features of this new release are:
  - Session rotation
  - Dynamic tracing
  - User and kernel space call stack capture
  - Support for arrays and bitwise binary operators in filters
  - Improved networking performance of the relay daemon

Read on for a short description of each of these features and
the links to this release!

A prettified version of this announcement is available here:
https://github.com/lttng/lttng-tools/releases/tag/v2.11.0-rc1

-- Session rotation --

The biggest feature of this release is the long-awaited session
rotation support. Session rotations now allow you to rotate an
ongoing tracing session much in the same way as you would rotate
logs.

The 'lttng rotate' command archives the current trace chunk of the
current tracing session. Once a trace chunk is archived, LTTng
does not manage it anymore: you can read it, modify it, move it,
or remove it.

Because a rotation causes the tracing session’s current sub-buffers
to be flushed, archived trace chunks are never redundant, that is, they
do not overlap over time like snapshots can.

Once a rotation is complete, offline analyses can be performed on
the resulting trace, much like in 'normal' mode. However, the big
advantage is that this can be done without interrupting tracing, and
without being limited to tools which implement the "live" protocol.

Moreover, session rotations can be scheduled on a time or size
basis. The following commands give an idea of how this functionality
can be used:

$ lttng enable-rotation --timer 30s

A rotation schedule is set so that an automatic rotation occurs at
least every 30 seconds.

$ lttng enable-rotation --size 100M

A the rotation schedule is set so that an automatic rotation occurs
every time the total size of the flushed part of the current trace
chunk is at least 100 MiB.

Conversely, session rotation schedules can be deactivated using the
'lttng disable-rotation' command.

Finally, it is also possible to use the Notification API to notify
external tools when a rotation has been completed. This allows
arbitrary actions to be taken on a trace archive when a scheduled
rotation is completed. Such actions can include compressing, analyzing,
or transferring a trace chunk, for example.

-- Dynamic tracing --

This release also marks the introduction of another feature we
have been eager to introduce for a long time: dynamic tracing.

As of this release, it is possible to dynamically instrument
function entries of applications and shared libraries, as well as
DTrace-style SDT probes.

The feature relies on the Linux kernel's uprobe facilities. As such,
the kernel tracer is needed to benefit from this functionality.

Here is an example instrumenting the libc's 'malloc' function
which will produce an event named 'libc_malloc' on every
invocation of malloc() by any application linked on libc:

$ lttng enable-event --kernel \
                     --userspace-probe /usr/lib/libc-2.28.so:malloc \
                     libc_malloc

SDT probes are tracepoints distributed in multiple libraries and
applications. This implementation supports tracing of SDT probes that
are NOT guarded by a semaphore.

The syntax to enable an SDT tracepoint userspace probe follows:

$ lttng enable-event --kernel \
                     --userspace-probe=sdt:/path/to/executable:provider:probe

Note that, for the time being, no payload is extracted (function
arguments or SDT probe parameters).

-- User and kernel space call stack capture --

Two new contexts have been introduced for kernel channels:
callstack-kernel and callstack-user.

This allows the sampling of both the kernel and user space call stacks
when a kernel event occurs.

For example, it is now possible to capture both user and kernel call
stacks whenever a process issues a 'read()' system call using the
following commands:

$ lttng create my_session
$ lttng enable-event --kernel --syscall read
$ lttng add-context --kernel --type callstack-kernel \
                             --type callstack-user
$ lttng start

Note that it is not always possible for the tracer to reliably sample
the user space call stack. For instance, it is not possible to sample
the call stack of user space code compiled with -fomit-frame pointer.
In such a case, the tracing is not affected, but the sampled user
space call stack may only contain the top-most address of the user space
call stack.

-- Support for arrays and bitwise binary operators in filters --

The filtering expression language and interpreter have been extended
to support new features:
  - Bracket and dot notations to get nested fields
  - New bitwise operators (>>, <<, &, ^, and |)

These improvements make it possible to filter on the content of
structures, arrays and sequences. The objective behind this work was
to make it easier to filter events which contain binary payloads
such as network headers.

As such, it is now possible to use filter expressions of the form:
"payload[2] & 0xff7 == 0x240"

Also, the documentation of filter expressions has been expanded in the
lttng-enable-event(1) man page to include more examples, a table of
operator precedence and type promotion rules.

-- Improved networking performance of the relay daemon --

While not a feature per se, a substantial amount of work went into
mitigating performance degradations which were reported when a large
number of targets streaming to a given relay daemon were suddenly
experiencing connectivity issues.

In those scenarios, the relay daemon could timeout on a number of
network operations, making it unresponsive and unable to process
incoming traffic for a long moment.

In order to make it more resilient in those situations, the relay's
communication with consumer daemons (i.e. traced targets) has been
made entirely asynchronous, thus mitigating issues caused by
misbehaving consumers. This is the first of many improvements to the
relay daemon's networking code.  More changes, requiring protocol
changes, are scheduled to make it in by the next 2.12 release.

Links
---

Project website: https://lttng.org

Download link:
https://lttng.org/files/lttng-tools/lttng-tools-2.11.0-rc1.tar.bz2

GPG sig:
https://lttng.org/files/lttng-tools/lttng-tools-2.11.0-rc1.tar.bz2.asc

Jérémie

-- 
Jérémie Galarneau
EfficiOS Inc.
http://www.efficios.com