[go: nahoru, domu]

Tips for debugging on Linux

This page is for Chromium-specific debugging tips; learning how to run gdb is out of scope.

Symbolized stack trace

The sandbox can interfere with the internal symbolizer. Use --no-sandbox (but keep this temporary) or an external symbolizer (see tools/valgrind/asan/asan_symbolize.py).

Generally, do not use --no-sandbox on waterfall bots, sandbox testing is needed. Talk to security@chromium.org.

GDB

GDB-7.7 is required in order to debug Chrome on Linux.

Any prior version will fail to resolve symbols or segfault.

Setup

Build setup

In your build set the GN build variable symbol_level = 2 for interactive debugging. (symbol_level = 1 only provides backtrace information). And while release-mode debugging is possible, things will be much easier in a debug build. Set your build args with gn args out/<your_dir> (substituting your build directory), and set:

is_debug = true
symbol_level = 2

GDB setup

The Chrome build requires some GDB configuration for it to be able to find source files. See gdbinit to configure GDB. There is a similar process for LLDB.

Basic browser process debugging

gdb -tui -ex=r --args out/Debug/chrome --disable-seccomp-sandbox \
    http://google.com

Allowing attaching to foreign processes

On distributions that use the Yama LSM (that includes Ubuntu and Chrome OS), process A can attach to process B only if A is an ancestor of B.

You will probably want to disable this feature by using

echo 0 | sudo tee /proc/sys/kernel/yama/ptrace_scope

If you don‘t you’ll get an error message such as “Could not attach to process”.

Note that you'll also probably want to use --no-sandbox, as explained below.

Multiprocess Tricks

Getting renderer subprocesses into gdb

Since Chromium itself spawns the renderers, it can be tricky to grab a particular with gdb. This command does the trick:

chrome --no-sandbox --renderer-cmd-prefix='xterm -title renderer -e gdb --args'

The --no-sandbox flag is needed because otherwise the seccomp sandbox will kill the renderer process on startup, or the setuid sandbox will prevent xterm‘s execution. The “xterm” is necessary or gdb will run in the current terminal, which can get particularly confusing since it’s running in the background, and if you‘re also running the main process in gdb, won’t work at all (the two instances will fight over the terminal). To auto-start the renderers in the debugger, send the “run” command to the debugger:

chrome --no-sandbox --renderer-cmd-prefix='xterm -title renderer -e gdb \
    -ex run --args'

If you're using Emacs and M-x gdb, you can do

chrome "--renderer-cmd-prefix=gdb --args"
Note: using the --renderer-cmd-prefix option bypasses the zygote launcher, so the renderers won‘t be sandboxed. It is generally not an issue, except when you are trying to debug interactions with the sandbox. If that’s what you are doing, you will need to attach your debugger to a running renderer process (see below).

You may also want to pass --disable-hang-monitor to suppress the hang monitor, which is rather annoying.

You can also use --renderer-startup-dialog and attach to the process in order to debug the renderer code. Go to https://www.chromium.org/blink/getting-started-with-blink-debugging for more information on how this can be done.

For utilities you can use --utility-startup-dialog to have all utilities prompt, or --utility-startup-dialog=data_decoder.mojom.DataDecoderService to debug only a particular service type.

Choosing which renderers to debug

If you are starting multiple renderers then the above means that multiple gdb's start and fight over the console. Instead, you can set the prefix to point to this shell script:

#!/bin/sh

echo "**** Child $$ starting: y to debug"
read input
if [ "$input" = "y" ] ; then
  gdb --args $*
else
  $*
fi

Choosing renderer to debug by URL

In most cases you'll want to debug the renderer which is loading a particular site. If you want a script which will automatically debug the renderer which has visited a given target URL and continue all other renderers, you can use the following:

./third_party/blink/tools/debug_renderer out/Default/content_shell https://example.domain/path

The script also supports specifying a different URL than the navigation URL. This is useful when the renderer you want to debug is not the top frame but one of the subframes on the page. For example, you could debug a particular subframe on a page with:

./third_party/blink/tools/debug_renderer -d https://subframe.url/path out/Default/content_shell https://example.domain/path

However, if you need more fine-grained control over which renderers to debug you can run chrome or content_shell directly with the --wait-for-debugger-on-navigation flag which will pause each renderer at the point of navigation (when the URL is known).

This will result in a series of lines such as the following in the output:

...:content_switches_internal.cc(119)] Renderer url="https://example.domain/path" (PID) paused waiting for debugger to attach. Send SIGUSR1 to unpause.

You can signal the renderers you aren't interested in to continue running with:

kill -s SIGUSR1 <pid>

And debug the renderer you are interested in debugging with:

gdb -p <pid>

Debugging run_web_tests.py renderers

The debug_renderer script can also be used to debug the renderer running a web test. To do so, simply call run_{web,wpt}_tests.py from debug_renderer with all of the standard arguments for run_{web,wpt}_tests.py. For example:

./third_party/blink/tools/debug_renderer ./third_party/blink/tools/run_web_tests.py [run_web_test args]

Selective breakpoints

When debugging both the browser and renderer process, you might want to have separate set of breakpoints to hit. You can use gdb's command files to accomplish this by putting breakpoints in separate files and instructing gdb to load them.

gdb -x ~/debug/browser --args chrome --no-sandbox --disable-hang-monitor \
    --renderer-cmd-prefix='xterm -title renderer -e gdb -x ~/debug/renderer \
    --args '

Also, instead of running gdb, you can use the script above, which let's you select which renderer process to debug. Note: you might need to use the full path to the script and avoid $HOME or ~/.

Connecting to a running renderer

Usually ps aux | grep chrome will not give very helpful output. Try pstree -p | grep chrome to get something like

        |                      |-bash(21969)---chrome(672)-+-chrome(694)
        |                      |                           |-chrome(695)---chrome(696)-+-{chrome}(697)
        |                      |                           |                           \-{chrome}(709)
        |                      |                           |-{chrome}(675)
        |                      |                           |-{chrome}(678)
        |                      |                           |-{chrome}(679)
        |                      |                           |-{chrome}(680)
        |                      |                           |-{chrome}(681)
        |                      |                           |-{chrome}(682)
        |                      |                           |-{chrome}(684)
        |                      |                           |-{chrome}(685)
        |                      |                           |-{chrome}(705)
        |                      |                           \-{chrome}(717)

Most of those are threads. In this case the browser process would be 672 and the (sole) renderer process is 696. You can use gdb -p 696 to attach. Alternatively, you might find out the process ID from Chrome's built-in Task Manager (under the Tools menu). Right-click on the Task Manager, and enable “Process ID” in the list of columns.

Note: by default, sandboxed processes can't be attached by a debugger. To be able to do so, you will need to pass the --allow-sandbox-debugging option.

If the problem only occurs with the seccomp sandbox enabled (and the previous tricks don‘t help), you could try enabling core-dumps (see the Core files section). That would allow you to get a backtrace and see some local variables, though you won’t be able to step through the running program.

Note: If you‘re interested in debugging LinuxSandboxIPC process, you can attach to 694 in the above diagram. The LinuxSandboxIPC process has the same command line flag as the browser process so that it’s easy to identify it if you run pstree -pa.

Getting GPU subprocesses into gdb

Use --gpu-launcher flag instead of --renderer-cmd-prefix in the instructions for renderer above.

Getting browser_tests launched browsers into gdb

Use environment variable BROWSER_WRAPPER instead of --renderer-cmd-prefix switch in the instructions above.

Example:

BROWSER_WRAPPER='xterm -title renderer -e gdb --eval-command=run \
    --eval-command=quit --args' out/Debug/browser_tests --gtest_filter=Print

Plugin Processes

Same strategies as renderers above, but the flag is called --plugin-launcher:

chrome --plugin-launcher='xterm -e gdb --args'
Note: For now, this does not currently apply to PPAPI plugins because they currently run in the renderer process.

Single-Process mode

Depending on whether it‘s relevant to the problem, it’s often easier to just run in “single process” mode where the renderer threads are in-process. Then you can just run gdb on the main process.

gdb --args chrome --single-process

Currently, the --disable-gpu flag is also required, as there are known crashes that occur under TextureImageTransportSurface without it. The crash described in https://crbug.com/361689 can also sometimes occur, but that crash can be continued from without harm.

Note that for technical reasons plugins cannot be in-process, so --single-process only puts the renderers in the browser process. The flag is still useful for debugging plugins (since it‘s only two processes instead of three) but you’ll still need to use --plugin-launcher or another approach.

Printing Chromium types

gdb 7 lets us use Python to write pretty-printers for Chromium types. See gdbinit to enable pretty-printing of Chromium types. This will import Blink pretty-printers as well.

Pretty printers for std types shouldn‘t be necessary in gdb 7, but they’re provided here in case you're using an older gdb. Put the following into ~/.gdbinit:

# Print a C++ string.
define ps
  print $arg0.c_str()
end

# Print a C++ wstring or wchar_t*.
define pws
  printf "\""
  set $c = (wchar_t*)$arg0
  while ( *$c )
    if ( *$c > 0x7f )
      printf "[%x]", *$c
    else
      printf "%c", *$c
    end
    set $c++
  end
  printf "\"\n"
end

More STL GDB macros

JsDbg -- visualize data structures in the browser

JsDbg is a debugger plugin to display various Chrome data structures in a browser window, such as the accessibility tree, layout object tree, DOM tree, and others. Installation instructions are here, and see here for screenshots and an introduction.

For Googlers, please see go/jsdbg for installation instructions.

Time travel debugging with rr

You can use rr for time travel debugging, so you can also step or execute backwards. This works by first recording a trace and then debugging based on that.

You need an up-to-date version of rr, since rr is frequently updated to support new parts of the Linux system call API surface that Chromium uses. If you have any issues with the latest release version, try compiling rr from source.

Once installed, you can use it like this:

rr record out/Debug/content_shell --single-process
rr replay
(rr) c
(rr) break blink::NGBlockNode::Layout
(rr) rc # reverse-continue to the last Layout call
(rr) jsdbg # run JsDbg as described above to find the interesting object
(rr) watch -l box_->frame_rect_.size_.width_.value_
(rr) rc # reverse-continue to the last time the width was changed
(rr) rn # reverse-next to the previous line
(rr) reverse-fin # run to where this function was called from

You can debug multi-process chrome using rr -f [PID] for processes fork()ed from a zygote process without exec, which includes renderer processes, or rr -p [PID] for other processes. To find the process id you can either run rr ps after recording, or for renderer processes use --vmodule=render_frame_impl=1 which will log a message on navigations. Example:

$ rr record out/Debug/content_shell --disable-hang-monitor --vmodule=render_frame_impl=1 https://www.google.com/
rr: Saving execution to trace directory `...'.
...
[128515:128515:0320/164124.768687:VERBOSE1:render_frame_impl.cc(4244)] Committed provisional load: https://www.google.com/

From the log message we can see that the site was loaded into process 128515 and can set a breakpoint for when that process is forked.

rr replay -f 128515

If you want to call debugging functions from gdb that use LOG(), then those functions need to disable the printing of timestamps using SetLogItems. See LayoutObject::ShowLayoutObject() for an example of this, and issue 2829 for why it is needed.

If rr doesn‘t work correctly, the rr developers are generally quite responsive to bug reports, especially ones that have enough information so that they don’t have to build Chromium.

See Also:

Graphical Debugging Aid for Chromium Views

The following link describes a tool that can be used on Linux, Windows and Mac under GDB.

graphical_debugging_aid_chromium_views

Faster startup

Use the gdb-add-index script (e.g. build/gdb-add-index out/Debug/browser_tests)

Only makes sense if you run the binary multiple times or maybe if you use the component build since most .so files won't require reindexing on a rebuild.

See https://groups.google.com/a/chromium.org/forum/#!searchin/chromium-dev/gdb-add-index/chromium-dev/ELRuj1BDCL4/5Ki4LGx41CcJ for more info.

You can improve GDB load time significantly at the cost of link time by not splitting symbols from the object files. In GN, set use_debug_fission=false in your “gn args”.

Core files

ulimit -c unlimited should cause all Chrome processes (run from that shell) to dump cores, with the possible exception of some sandboxed processes.

Some sandboxed subprocesses might not dump cores unless you pass the --allow-sandbox-debugging flag.

If the problem is a freeze rather than a crash, you may be able to trigger a core-dump by sending SIGABRT to the relevant process:

kill -6 [process id]

Breakpad minidump files

See minidump_to_core.md

Running Tests

Many of our tests bring up windows on screen. This can be annoying (they steal your focus) and hard to debug (they receive extra events as you mouse over them). Instead, use Xvfb or Xephyr to run a nested X session to debug them, as outlined on testing/web_tests_linux.md.

Browser tests

By default the browser_tests forks a new browser for each test. To debug the browser side of a single test, use a command like

gdb --args out/Debug/browser_tests --single-process-tests --gtest_filter=MyTestName

note the use of single-process-tests -- this makes the test harness and browser process share the outermost process.

The switch --gtest_break_on_failure can also be useful to automatically stop debugger upon ASSERT or EXPECT failures.

To debug a renderer process in this case, use the tips above about renderers.

Web tests

See testing/web_tests_linux.md for some tips. In particular, note that it‘s possible to debug a web test via sshing to a Linux box; you don’t need anything on screen if you use Xvfb.

UI tests

UI tests are run in forked browsers. Unlike browser tests, you cannot do any single process tricks here to debug the browser. See below about BROWSER_WRAPPER.

To pass flags to the browser, use a command line like --extra-chrome-flags="--foo --bar".

Timeouts

UI tests have a confusing array of timeouts in place. (Pawel is working on reducing the number of timeouts.) To disable them while you debug, set the timeout flags to a large value:

  • --test-timeout=100000000
  • --ui-test-action-timeout=100000000
  • --ui-test-terminate-timeout=100000000

To replicate Window Manager setup on the bots

Chromium try bots and main waterfall's bots run tests under Xvfb&openbox combination. Xvfb is an X11 server that redirects the graphical output to the memory, and openbox is a simple window manager that is running on top of Xvfb. The behavior of openbox is markedly different when it comes to focus management and other window tasks, so test that runs fine locally may fail or be flaky on try bots. To run the tests on a local machine as on a bot, follow these steps:

Make sure you have openbox:

apt-get install openbox

Start Xvfb and openbox on a particular display:

Xvfb :6.0 -screen 0 1280x1024x24 & DISPLAY=:6.0 openbox &

Run your tests with graphics output redirected to that display:

DISPLAY=:6.0 out/Debug/browser_tests --gtest_filter="MyBrowserTest.MyActivateWindowTest"

You can look at a snapshot of the output by:

xwd -display :6.0 -root | xwud

Alternatively, you can use testing/xvfb.py to set up your environment for you:

testing/xvfb.py out/Debug/browser_tests \
    --gtest_filter="MyBrowserTest.MyActivateWindowTest"

BROWSER_WRAPPER

You can also get the browser under a debugger by setting the BROWSER_WRAPPER environment variable. (You can use this for browser_tests too, but see above for discussion of a simpler way.)

BROWSER_WRAPPER='xterm -e gdb --args' out/Debug/browser_tests

Replicating try bot Slowness

Try bots are pretty stressed, and can sometimes expose timing issues you can't normally reproduce locally.

You can simulate this by shutting down all but one of the CPUs (http://www.cyberciti.biz/faq/debian-rhel-centos-redhat-suse-hotplug-cpu/) and running a CPU loading tool (e.g., http://www.devin.com/lookbusy/). Now run your test. It will run slowly, but any flakiness found by the try bot should replicate locally now - and often nearly 100% of the time.

Logging

Seeing all LOG(foo) messages

Default log level hides LOG(INFO). Run with --log-level=0 and --enable-logging=stderr flags.

Newer versions of Chromium with VLOG may need --v=1 too. For more VLOG tips, see the chromium-dev thread.

Seeing IPC debug messages

Run with CHROME_IPC_LOGGING=1 eg.

CHROME_IPC_LOGGING=1 out/Debug/chrome

or within gdb:

set environment CHROME_IPC_LOGGING 1

If some messages show as unknown, check if the list of IPC message headers in chrome/common/logging_chrome.cc is up to date. In case this file reference goes out of date, try looking for usage of macros like IPC_MESSAGE_LOG_ENABLED or IPC_MESSAGE_MACROS_LOG_ENABLED.

Profiling

See https://sites.google.com/a/chromium.org/dev/developers/profiling-chromium-and-webkit and Linux Profiling.

i18n

We obey your system locale. Try something like:

LANG=ja_JP.UTF-8 out/Debug/chrome

If this doesn‘t work, make sure that the LANGUAGE, LC_ALL and LC_MESSAGE environment variables aren’t set -- they have higher priority than LANG in the order listed. Alternatively, just do this:

LANGUAGE=fr out/Debug/chrome

Note that because we use GTK, some locale data comes from the system -- for example, file save boxes and whether the current language is considered RTL. Without all the language data available, Chrome will use a mixture of your system language and the language you run Chrome in.

Here's how to install the Arabic (ar) and Hebrew (he) language packs:

sudo apt-get install language-pack-ar language-pack-he \
    language-pack-gnome-ar language-pack-gnome-he

Note that the --lang flag does not work properly for this.

On non-Debian systems, you need the gtk30.mo files. (Please update these docs with the appropriate instructions if you know what they are.)

Breakpad

See the last section of Linux Crash Dumping.

Drag and Drop

If you break in a debugger during a drag, Chrome will have grabbed your mouse and keyboard so you won't be able to interact with the debugger! To work around this, run via Xephyr. Instructions for how to use Xephyr are on the Running web tests on Linux page.

Tracking Down Bugs

Isolating Regressions

Old builds are archived here: https://build.chromium.org/buildbot/snapshots/chromium-rel-linux/ (TODO: does not exist).

tools/bisect-builds.py in the tree automates bisecting through the archived builds. Despite a computer science education, I am still amazed how quickly binary search will find its target.

Screen recording for bug reports

sudo apt-get install gtk-recordmydesktop

Version-specific issues

Google Chrome

Google Chrome binaries don't include symbols. Googlers can read where to get symbols from the Google-internal wiki.

Ubuntu Chromium

Since we don‘t build the Ubuntu packages (Ubuntu does) we can’t get useful backtraces from them. Direct users to https://wiki.ubuntu.com/Chromium/Debugging

Fedora's Chromium

Like Ubuntu, but direct users to https://fedoraproject.org/wiki/TomCallaway/Chromium_Debug

Xlib

If you're trying to track down X errors like:

The program 'chrome' received an X Window System error.
This probably reflects a bug in the program.
The error was 'BadDrawable (invalid Pixmap or Window parameter)'.

Some strategies are:

  • pass --sync on the command line to make all X calls synchronous
  • run chrome via xtrace
  • turn on IPC debugging (see above section)

Window Managers

To test on various window managers, you can use a nested X server like Xephyr. Instructions for how to use Xephyr are on the Running web tests on Linux page.

If you need to test something with hardware accelerated compositing (e.g., compiz), you can use Xgl (sudo apt-get install xserver-xgl). E.g.:

Xgl :1 -ac -accel glx:pbuffer -accel xv:pbuffer -screen 1024x768

Mozilla Tips

https://developer.mozilla.org/en/Debugging_Mozilla_on_Linux_FAQ

Google Chrome Symbol Files

Symbols for Google Chrome's official builds are available from https://edgedl.me.gvt1.com/chrome/linux/symbols/google-chrome-debug-info-linux64-${VERSION}.zip where ${VERSION} is any version of Google Chrome that has recently been served to Stable, Beta, or Unstable (Dev) channels on Linux.