5. Trace Library

5.1. Overview

Tracing is a technique used to understand what goes on in a running software system. The software used for tracing is called a tracer, which is conceptually similar to a tape recorder. When recording, specific instrumentation points placed in the software source code generate events that are saved on a giant tape: a trace file. The trace file then later can be opened in trace viewers to visualize and analyze the trace events with timestamps and multi-core views. Such a mechanism will be useful for resolving a wide range of problems such as multi-core synchronization issues, latency measurements, finding out the post analysis information like CPU idle time, etc that would otherwise be extremely challenging to get.

Tracing is often compared to logging. However, tracers and loggers are two different tools, serving two different purposes. Tracers are designed to record much lower-level events that occur much more frequently than log messages, often in the range of thousands per second, with very little execution overhead. Logging is more appropriate for a very high-level analysis of less frequent events: user accesses, exceptional conditions (errors and warnings, for example), database transactions, instant messaging communications, and such. Simply put, logging is one of the many use cases that can be satisfied with tracing.

5.2. DPDK tracing library features

  • A framework to add tracepoints in control and fast path APIs with minimum impact on performance. Typical trace overhead is ~20 cycles and instrumentation overhead is 1 cycle.
  • Enable and disable the tracepoints at runtime.
  • Save the trace buffer to the filesystem at any point in time.
  • Support overwrite and discard trace mode operations.
  • String-based tracepoint object lookup.
  • Enable and disable a set of tracepoints based on regular expression and/or globbing.
  • Generate trace in Common Trace Format (CTF). CTF is an open-source trace format and is compatible with LTTng. For detailed information, refer to Common Trace Format.

5.3. How to add a tracepoint?

This section steps you through the details of adding a simple tracepoint.

5.3.1. Create the tracepoint header file

#include <rte_trace_point.h>

RTE_TRACE_POINT(
       app_trace_string,
       RTE_TRACE_POINT_ARGS(const char *str),
       rte_trace_point_emit_string(str);
)

The above macro creates app_trace_string tracepoint. The user can choose any name for the tracepoint. However, when adding a tracepoint in the DPDK library, the rte_<library_name>_trace_[<domain>_]<name> naming convention must be followed. The examples are rte_eal_trace_generic_str, rte_mempool_trace_create.

The RTE_TRACE_POINT macro expands from above definition as the following function template:

static __rte_always_inline void
app_trace_string(const char *str)
{
        /* Trace subsystem hooks */
        ...
        rte_trace_point_emit_string(str);
}

The consumer of this tracepoint can invoke app_trace_string(const char *str) to emit the trace event to the trace buffer.

5.3.2. Register the tracepoint

#include <rte_trace_point_register.h>

#include <my_tracepoint.h>

RTE_TRACE_POINT_DEFINE(app_trace_string);

RTE_INIT(app_trace_init)
{
      RTE_TRACE_POINT_REGISTER(app_trace_string, app.trace.string);
}

The above code snippet registers the app_trace_string tracepoint to trace library. Here, the my_tracepoint.h is the header file that the user created in the first step Create the tracepoint header file.

The second argument for the RTE_TRACE_POINT_REGISTER is the name for the tracepoint. This string will be used for tracepoint lookup or regular expression and/or glob based tracepoint operations. There is no requirement for the tracepoint function and its name to be similar. However, it is recommended to have a similar name for a better naming convention.

The user must register the tracepoint before the rte_eal_init invocation. The user can use the RTE_INIT construction scheme to achieve this.

Note

The rte_trace_point_register.h header must be included before any inclusion of the rte_trace_point.h header.

Note

The RTE_TRACE_POINT_DEFINE defines the placeholder for the rte_trace_point_t tracepoint object. The user must export a __<trace_function_name> symbol in the library .map file for this tracepoint to be used out of the library, in shared builds. For example, __app_trace_string will be the exported symbol in the above example.

5.4. Fast path tracepoint

In order to avoid performance impact in fast path code, the library introduced RTE_TRACE_POINT_FP. When adding the tracepoint in fast path code, the user must use RTE_TRACE_POINT_FP instead of RTE_TRACE_POINT.

RTE_TRACE_POINT_FP is compiled out by default and it can be enabled using CONFIG_RTE_ENABLE_TRACE_FP configuration parameter. The enable_trace_fp option shall be used for the same for meson build.

5.5. Event record mode

Event record mode is an attribute of trace buffers. Trace library exposes the following modes:

Overwrite
When the trace buffer is full, new trace events overwrites the existing captured events in the trace buffer.
Discard
When the trace buffer is full, new trace events will be discarded.

The mode can be configured either using EAL command line parameter --trace-mode on application boot up or use rte_trace_mode_set() API to configure at runtime.

5.6. Trace file location

On rte_trace_save() or rte_eal_cleanup() invocation, the library saves the trace buffers to the filesystem. By default, the trace files are stored in $HOME/dpdk-traces/rte-yyyy-mm-dd-[AP]M-hh-mm-ss/. It can be overridden by the --trace-dir=<directory path> EAL command line option.

For more information, refer to EAL parameters for trace EAL command line options.

5.7. View and analyze the recorded events

Once the trace directory is available, the user can view/inspect the recorded events.

There are many tools you can use to read DPDK traces:

1. babeltrace is a command-line utility that converts trace formats; it supports the format that DPDK trace library produces, CTF, as well as a basic text output that can be grep’ed. The babeltrace command is part of the Open Source Babeltrace project.

2. Trace Compass is a graphical user interface for viewing and analyzing any type of logs or traces, including DPDK traces.

5.7.1. Use the babeltrace command-line tool

The simplest way to list all the recorded events of a trace is to pass its path to babeltrace with no options:

babeltrace </path-to-trace-events/rte-yyyy-mm-dd-[AP]M-hh-mm-ss/>

babeltrace finds all traces recursively within the given path and prints all their events, merging them in chronological order.

You can pipe the output of the babeltrace into a tool like grep(1) for further filtering. Below example grep the events for ethdev only:

babeltrace /tmp/my-dpdk-trace | grep ethdev

You can pipe the output of babeltrace into a tool like wc(1) to count the recorded events. Below example count the number of ethdev events:

babeltrace /tmp/my-dpdk-trace | grep ethdev | wc --lines

5.7.2. Use the tracecompass GUI tool

Tracecompass is another tool to view/analyze the DPDK traces which gives a graphical view of events. Like babeltrace, tracecompass also provides an interface to search for a particular event. To use tracecompass, following are the minimum required steps:

  • Install tracecompass to the localhost. Variants are available for Linux, Windows, and OS-X.
  • Launch tracecompass which will open a graphical window with trace management interfaces.
  • Open a trace using File->Open Trace option and select metadata file which is to be viewed/analyzed.

For more details, refer Trace Compass.

5.8. Quick start

This section steps you through the details of generating trace and viewing it.

  • Start the dpdk-test:

    echo "quit" | ./build/app/test/dpdk-test --no-huge --trace=.*

  • View the traces with babeltrace viewer:

    babeltrace $HOME/dpdk-traces/rte-yyyy-mm-dd-[AP]M-hh-mm-ss/

5.9. Implementation details

As DPDK trace library is designed to generate traces that uses Common Trace Format (CTF). CTF specification consists of the following units to create a trace.

  • Stream Sequence of packets.
  • Packet Header and one or more events.
  • Event Header and payload.

For detailed information, refer to Common Trace Format.

The implementation details broadly divided into the following areas:

5.9.1. Trace metadata creation

Based on the CTF specification, one of a CTF trace’s streams is mandatory: the metadata stream. It contains exactly what you would expect: data about the trace itself. The metadata stream contains a textual description of the binary layouts of all the other streams.

This description is written using the Trace Stream Description Language (TSDL), a declarative language that exists only in the realm of CTF. The purpose of the metadata stream is to make CTF readers know how to parse a trace’s binary streams of events without CTF specifying any fixed layout. The only stream layout known in advance is, in fact, the metadata stream’s one.

The internal trace_metadata_create() function generates the metadata.

5.9.2. Trace memory

The trace memory will be allocated through an internal function __rte_trace_mem_per_thread_alloc(). The trace memory will be allocated per thread to enable lock less trace-emit function. The memory for the trace memory for DPDK lcores will be allocated on rte_eal_init() if the trace is enabled through a EAL option. For non DPDK threads, on the first trace emission, the memory will be allocated.

5.9.3. Trace memory layout

Table 5.1 Trace memory layout.
packet.header
packet.context
trace 0 header
trace 0 payload
trace 1 header
trace 1 payload
trace N header
trace N payload

5.9.3.1. packet.header

Table 5.2 Packet header layout.
uint32_t magic
rte_uuid_t uuid

5.9.3.2. packet.context

Table 5.3 Packet context layout.
uint32_t thread_id
char thread_name[32]

5.9.3.3. trace.header

Table 5.4 Trace header layout.
event_id [63:48]
timestamp [47:0]

The trace header is 64 bits, it consists of 48 bits of timestamp and 16 bits event ID.

The packet.header and packet.context will be written in the slow path at the time of trace memory creation. The trace.header and trace payload will be emitted when the tracepoint function is invoked.