documentation/dev-manual/dev-manual-common-tasks.xml: Moved dbg and profile

Moved the sections on remote debugging and using the profile out of
the YP Reference Manual and into the YP Development Manual.  These are
now two sections in the "Common Tasks" chapter.

(From yocto-docs rev: 42640bf69aa05c56eb112c10b6d7b96069173785)

Signed-off-by: Scott Rifenbark <scott.m.rifenbark@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
This commit is contained in:
Scott Rifenbark 2012-06-18 09:23:29 -07:00 committed by Richard Purdie
parent 6b8922dbcf
commit b2f4ddd6b4
1 changed files with 478 additions and 0 deletions

View File

@ -1726,6 +1726,484 @@ so that there are some definite steps on how to do this. I need more detail her
process.
</para>
</section>
<section id="platdev-gdb-remotedebug">
<title>Debugging With the GNU Project Debugger (GDB) Remotely</title>
<para>
GDB allows you to examine running programs, which in turn help you to understand and fix problems.
It also allows you to perform post-mortem style analysis of program crashes.
GDB is available as a package within the Yocto Project and by default is
installed in sdk images.
See the "<ulink url='&YOCTO_DOCS_REF_URL;#ref-images'>Reference: Images</ulink>" appendix
in the Yocto Project Reference Manual for a description of these images.
You can find information on GDB at <ulink url="http://sourceware.org/gdb/"/>.
</para>
<tip>
For best results, install <filename>-dbg</filename> packages for the applications
you are going to debug.
Doing so makes available extra debug symbols that give you more meaningful output.
</tip>
<para>
Sometimes, due to memory or disk space constraints, it is not possible
to use GDB directly on the remote target to debug applications.
These constraints arise because GDB needs to load the debugging information and the
binaries of the process being debugged.
Additionally, GDB needs to perform many computations to locate information such as function
names, variable names and values, stack traces and so forth - even before starting the
debugging process.
These extra computations place more load on the target system and can alter the
characteristics of the program being debugged.
</para>
<para>
To help get past the previously mentioned constraints, you can use Gdbserver.
Gdbserver runs on the remote target and does not load any debugging information
from the debugged process.
Instead, a GDB instance processes the debugging information that is run on a
remote computer - the host GDB.
The host GDB then sends control commands to Gdbserver to make it stop or start the debugged
program, as well as read or write memory regions of that debugged program.
All the debugging information loaded and processed as well
as all the heavy debugging is done by the host GDB.
Offloading these processes gives the Gdbserver running on the target a chance to remain
small and fast.
</para>
<para>
Because the host GDB is responsible for loading the debugging information and
for doing the necessary processing to make actual debugging happen, the
user has to make sure the host can access the unstripped binaries complete
with their debugging information and also be sure the target is compiled with no optimizations.
The host GDB must also have local access to all the libraries used by the
debugged program.
Because Gdbserver does not need any local debugging information, the binaries on
the remote target can remain stripped.
However, the binaries must also be compiled without optimization
so they match the host's binaries.
</para>
<para>
To remain consistent with GDB documentation and terminology, the binary being debugged
on the remote target machine is referred to as the "inferior" binary.
For documentation on GDB see the
<ulink url="http://sourceware.org/gdb/documentation/">GDB site</ulink>.
</para>
<section id="platdev-gdb-remotedebug-launch-gdbserver">
<title>Launching Gdbserver on the Target</title>
<para>
First, make sure Gdbserver is installed on the target.
If it is not, install the package <filename>gdbserver</filename>, which needs the
<filename>libthread-db1</filename> package.
</para>
<para>
As an example, to launch Gdbserver on the target and make it ready to "debug" a
program located at <filename>/path/to/inferior</filename>, connect
to the target and launch:
<literallayout class='monospaced'>
$ gdbserver localhost:2345 /path/to/inferior
</literallayout>
Gdbserver should now be listening on port 2345 for debugging
commands coming from a remote GDB process that is running on the host computer.
Communication between Gdbserver and the host GDB are done using TCP.
To use other communication protocols, please refer to the
<ulink url='http://www.gnu.org/software/gdb/'>Gdbserver documentation</ulink>.
</para>
</section>
<section id="platdev-gdb-remotedebug-launch-gdb">
<title>Launching GDB on the Host Computer</title>
<para>
Running GDB on the host computer takes a number of stages.
This section describes those stages.
</para>
<section id="platdev-gdb-remotedebug-launch-gdb-buildcross">
<title>Building the Cross-GDB Package</title>
<para>
A suitable GDB cross-binary is required that runs on your host computer but
also knows about the the ABI of the remote target.
You can get this binary from the the Yocto Project meta-toolchain.
Here is an example:
<literallayout class='monospaced'>
/usr/local/poky/eabi-glibc/arm/bin/arm-poky-linux-gnueabi-gdb
</literallayout>
where <filename>arm</filename> is the target architecture and
<filename>linux-gnueabi</filename> the target ABI.
</para>
<para>
Alternatively, the Yocto Project can build the <filename>gdb-cross</filename> binary.
Here is an example:
<literallayout class='monospaced'>
$ bitbake gdb-cross
</literallayout>
Once the binary is built, you can find it here:
<literallayout class='monospaced'>
tmp/sysroots/&lt;host-arch&gt;/usr/bin/&lt;target-abi&gt;-gdb
</literallayout>
</para>
</section>
<section id="platdev-gdb-remotedebug-launch-gdb-inferiorbins">
<title>Making the Inferior Binaries Available</title>
<para>
The inferior binary (complete with all debugging symbols) as well as any
libraries (and their debugging symbols) on which the inferior binary depends
need to be available.
There are a number of ways you can make these available.
</para>
<para>
Perhaps the easiest way is to have an 'sdk' image that corresponds to the plain
image installed on the device.
In the case of <filename>core-image-sato</filename>,
<filename>core-image-sato-sdk</filename> would contain suitable symbols.
Because the sdk images already have the debugging symbols installed, it is just a
question of expanding the archive to some location and then informing GDB.
</para>
<para>
Alternatively, Yocto Project can build a custom directory of files for a specific
debugging purpose by reusing its <filename>tmp/rootfs</filename> directory.
This directory contains the contents of the last built image.
This process assumes two things:
<itemizedlist>
<listitem><para>The image running on the target was the last image to
be built by the Yocto Project.</para></listitem>
<listitem><para>The package (<filename>foo</filename> in the following
example) that contains the inferior binary to be debugged has been built
without optimization and has debugging information available.</para></listitem>
</itemizedlist>
</para>
<para>
The following steps show how to build the custom directory of files:
<orderedlist>
<listitem><para>Install the package (<filename>foo</filename> in this case) to
<filename>tmp/rootfs</filename>:
<literallayout class='monospaced'>
$ tmp/sysroots/i686-linux/usr/bin/opkg-cl -f \
tmp/work/&lt;target-abi&gt;/core-image-sato-1.0-r0/temp/opkg.conf -o \
tmp/rootfs/ update
</literallayout></para></listitem>
<listitem><para>Install the debugging information:
<literallayout class='monospaced'>
$ tmp/sysroots/i686-linux/usr/bin/opkg-cl -f \
tmp/work/&lt;target-abi&gt;/core-image-sato-1.0-r0/temp/opkg.conf \
-o tmp/rootfs install foo
$ tmp/sysroots/i686-linux/usr/bin/opkg-cl -f \
tmp/work/&lt;target-abi&gt;/core-image-sato-1.0-r0/temp/opkg.conf \
-o tmp/rootfs install foo-dbg
</literallayout></para></listitem>
</orderedlist>
</para>
</section>
<section id="platdev-gdb-remotedebug-launch-gdb-launchhost">
<title>Launch the Host GDB</title>
<para>
To launch the host GDB, you run the <filename>cross-gdb</filename> binary and provide
the inferior binary as part of the command line.
For example, the following command form continues with the example used in
the previous section.
This command form loads the <filename>foo</filename> binary
as well as the debugging information:
<literallayout class='monospaced'>
$ &lt;target-abi&gt;-gdb rootfs/usr/bin/foo
</literallayout>
Once the GDB prompt appears, you must instruct GDB to load all the libraries
of the inferior binary from <filename>tmp/rootfs</filename> as follows:
<literallayout class='monospaced'>
$ set solib-absolute-prefix /path/to/tmp/rootfs
</literallayout>
The pathname <filename>/path/to/tmp/rootfs</filename> must either be
the absolute path to <filename>tmp/rootfs</filename> or the location at which
binaries with debugging information reside.
</para>
<para>
At this point you can have GDB connect to the Gdbserver that is running
on the remote target by using the following command form:
<literallayout class='monospaced'>
$ target remote remote-target-ip-address:2345
</literallayout>
The <filename>remote-target-ip-address</filename> is the IP address of the
remote target where the Gdbserver is running.
Port 2345 is the port on which the GDBSERVER is running.
</para>
</section>
<section id="platdev-gdb-remotedebug-launch-gdb-using">
<title>Using the Debugger</title>
<para>
You can now proceed with debugging as normal - as if you were debugging
on the local machine.
For example, to instruct GDB to break in the "main" function and then
continue with execution of the inferior binary use the following commands
from within GDB:
<literallayout class='monospaced'>
(gdb) break main
(gdb) continue
</literallayout>
</para>
<para>
For more information about using GDB, see the project's online documentation at
<ulink url="http://sourceware.org/gdb/download/onlinedocs/"/>.
</para>
</section>
</section>
</section>
<section id="platdev-oprofile">
<title>Profiling with OProfile</title>
<para>
<ulink url="http://oprofile.sourceforge.net/">OProfile</ulink> is a
statistical profiler well suited for finding performance
bottlenecks in both userspace software and in the kernel.
This profiler provides answers to questions like "Which functions does my application spend
the most time in when doing X?"
Because the Yocto Project is well integrated with OProfile, it makes profiling applications on target
hardware straightforward.
</para>
<para>
To use OProfile, you need an image that has OProfile installed.
The easiest way to do this is with <filename>tools-profile</filename> in the
<filename><ulink url='&YOCTO_DOCS_REF_URL;#var-IMAGE_FEATURES'>IMAGE_FEATURES</ulink></filename> variable.
You also need debugging symbols to be available on the system where the analysis
takes place.
You can gain access to the symbols by using <filename>dbg-pkgs</filename> in the
<filename>IMAGE_FEATURES</filename> variable or by
installing the appropriate <filename>-dbg</filename> packages.
</para>
<para>
For successful call graph analysis, the binaries must preserve the frame
pointer register and should also be compiled with the
<filename>-fno-omit-framepointer</filename> flag.
In the Yocto Project you can achieve this by setting the
<filename><ulink url='&YOCTO_DOCS_REF_URL;#var-SELECTED_OPTIMIZATION'>SELECTED_OPTIMIZATION</ulink></filename>
variable to
<filename>-fexpensive-optimizations -fno-omit-framepointer -frename-registers -O2</filename>.
You can also achieve it by setting the
<filename><ulink url='&YOCTO_DOCS_REF_URL;#var-DEBUG_BUILD'>DEBUG_BUILD</ulink></filename>
variable to "1" in the <filename>local.conf</filename> configuration file.
If you use the <filename>DEBUG_BUILD</filename> variable you will also add extra debug information
that can make the debug packages large.
</para>
<section id="platdev-oprofile-target">
<title>Profiling on the Target</title>
<para>
Using OProfile you can perform all the profiling work on the target device.
A simple OProfile session might look like the following:
</para>
<para>
<literallayout class='monospaced'>
# opcontrol --reset
# opcontrol --start --separate=lib --no-vmlinux -c 5
.
.
[do whatever is being profiled]
.
.
# opcontrol --stop
$ opreport -cl
</literallayout>
</para>
<para>
In this example, the <filename>reset</filename> command clears any previously profiled data.
The next command starts OProfile.
The options used when starting the profiler separate dynamic library data
within applications, disable kernel profiling, and enable callgraphing up to
five levels deep.
<note>
To profile the kernel, you would specify the
<filename>--vmlinux=/path/to/vmlinux</filename> option.
The <filename>vmlinux</filename> file is usually in the Yocto Project file's
<filename>/boot/</filename> directory and must match the running kernel.
</note>
</para>
<para>
After you perform your profiling tasks, the next command stops the profiler.
After that, you can view results with the <filename>opreport</filename> command with options
to see the separate library symbols and callgraph information.
</para>
<para>
Callgraphing logs information about time spent in functions and about a function's
calling function (parent) and called functions (children).
The higher the callgraphing depth, the more accurate the results.
However, higher depths also increase the logging overhead.
Consequently, you should take care when setting the callgraphing depth.
<note>
On ARM, binaries need to have the frame pointer enabled for callgraphing to work.
To accomplish this use the <filename>-fno-omit-framepointer</filename> option
with <filename>gcc</filename>.
</note>
</para>
<para>
For more information on using OProfile, see the OProfile
online documentation at
<ulink url="http://oprofile.sourceforge.net/docs/"/>.
</para>
</section>
<section id="platdev-oprofile-oprofileui">
<title>Using OProfileUI</title>
<para>
A graphical user interface for OProfile is also available.
You can download and build this interface from the Yocto Project at
<ulink url="&YOCTO_GIT_URL;/cgit.cgi/oprofileui/"></ulink>.
If the "tools-profile" image feature is selected, all necessary binaries
are installed onto the target device for OProfileUI interaction.
</para>
<para>
Even though the Yocto Project usually includes all needed patches on the target device, you
might find you need other OProfile patches for recent OProfileUI features.
If so, see the <ulink url='&YOCTO_GIT_URL;/cgit.cgi/oprofileui/tree/README'>
OProfileUI README</ulink> for the most recent information.
</para>
<section id="platdev-oprofile-oprofileui-online">
<title>Online Mode</title>
<para>
Using OProfile in online mode assumes a working network connection with the target
hardware.
With this connection, you just need to run "oprofile-server" on the device.
By default, OProfile listens on port 4224.
<note>
You can change the port using the <filename>--port</filename> command-line
option.
</note>
</para>
<para>
The client program is called <filename>oprofile-viewer</filename> and its UI is relatively
straightforward.
You access key functionality through the buttons on the toolbar, which
are duplicated in the menus.
Here are the buttons:
<itemizedlist>
<listitem><para><emphasis>Connect:</emphasis> Connects to the remote host.
You can also supply the IP address or hostname.</para></listitem>
<listitem><para><emphasis>Disconnect:</emphasis> Disconnects from the target.
</para></listitem>
<listitem><para><emphasis>Start:</emphasis> Starts profiling on the device.
</para></listitem>
<listitem><para><emphasis>Stop:</emphasis> Stops profiling on the device and
downloads the data to the local host.
Stopping the profiler generates the profile and displays it in the viewer.
</para></listitem>
<listitem><para><emphasis>Download:</emphasis> Downloads the data from the
target and generates the profile, which appears in the viewer.</para></listitem>
<listitem><para><emphasis>Reset:</emphasis> Resets the sample data on the device.
Resetting the data removes sample information collected from previous
sampling runs.
Be sure you reset the data if you do not want to include old sample information.
</para></listitem>
<listitem><para><emphasis>Save:</emphasis> Saves the data downloaded from the
target to another directory for later examination.</para></listitem>
<listitem><para><emphasis>Open:</emphasis> Loads previously saved data.
</para></listitem>
</itemizedlist>
</para>
<para>
The client downloads the complete 'profile archive' from
the target to the host for processing.
This archive is a directory that contains the sample data, the object files,
and the debug information for the object files.
The archive is then converted using the <filename>oparchconv</filename> script, which is
included in this distribution.
The script uses <filename>opimport</filename> to convert the archive from
the target to something that can be processed on the host.
</para>
<para>
Downloaded archives reside in the Yocto Project's build directory in
<filename>/tmp</filename> and are cleared up when they are no longer in use.
</para>
<para>
If you wish to perform kernel profiling, you need to be sure
a <filename>vmlinux</filename> file that matches the running kernel is available.
In the Yocto Project, that file is usually located in
<filename>/boot/vmlinux-KERNELVERSION</filename>, where
<filename>KERNEL-version</filename> is the version of the kernel.
The Yocto Project generates separate <filename>vmlinux</filename> packages for each kernel
it builds.
Thus, it should just be a question of making sure a matching package is
installed (e.g. <filename>opkg install kernel-vmlinux</filename>.
The files are automatically installed into development and profiling images
alongside OProfile.
A configuration option exists within the OProfileUI settings page that you can use to
enter the location of the <filename>vmlinux</filename> file.
</para>
<para>
Waiting for debug symbols to transfer from the device can be slow, and it
is not always necessary to actually have them on the device for OProfile use.
All that is needed is a copy of the filesystem with the debug symbols present
on the viewer system.
The "<link linkend='platdev-gdb-remotedebug-launch-gdb'>Launching GDB on the Host Computer</link>"
section covers how to create such a directory with
the Yocto Project and how to use the OProfileUI Settings dialog to specify the location.
If you specify the directory, it will be used when the file checksums
match those on the system you are profiling.
</para>
</section>
<section id="platdev-oprofile-oprofileui-offline">
<title>Offline Mode</title>
<para>
If network access to the target is unavailable, you can generate
an archive for processing in <filename>oprofile-viewer</filename> as follows:
<literallayout class='monospaced'>
# opcontrol --reset
# opcontrol --start --separate=lib --no-vmlinux -c 5
.
.
[do whatever is being profiled]
.
.
# opcontrol --stop
# oparchive -o my_archive
</literallayout>
</para>
<para>
In the above example, <filename>my_archive</filename> is the name of the
archive directory where you would like the profile archive to be kept.
After the directory is created, you can copy it to another host and load it
using <filename>oprofile-viewer</filename> open functionality.
If necessary, the archive is converted.
</para>
</section>
</section>
</section>
</chapter>
<!--