Debugging Chromium on Windows

 

First see get the code for checkout and build instructions.

Getting started

WinDBG

devenv /debugexe outDebugchrome.exe <options to Chromium can go here>

 

Note that the path to the binary must use backslashes and must include the ".exe" suffix or Visual Studio will open and do nothing. Note also that this assumes you have Visual Studio installed and have devenv.exe on your path. You can open a Visual Studio developer command prompt or run this command to add it to your path:

 

"C:Program Files (x86)Microsoft Visual Studio2019ProfessionalVCAuxiliaryBuildvcvarsall.bat"

 

Visual Studio 2017 is not recommended for debugging of Chromium - use a newer version for best performance and stability.

 

symbol_level=2 is the default on Windows and gives full debugging information with types, locals, globals, function names, and source/line information. symbol_level=1 creates smaller PDBs with just function names, and source/line information - source-level debugging is still supported (new from June 2019).

Profiles

It's a good idea to use a different Chrome profile for your debugging. If you are debugging Google Chrome branded builds, or use a Chromium build as your primary browser, the profiles can collide so you can't run both at once, and your stable browser might see profile versions from the future (Google Chrome and Chromium use different profile directories by default so won't collide). Use the command-line option:

 

--user-data-dir=c: mpmy_debug_profile (replace the path as necessary)

 

Using the IDE, go to the Debugging tab of the properties of the chrome project, and set the Command Arguments.

Chrome debug log

--enable-logging --v=1 command-line flags at startup.  file in the  directory. Release builds place the file in the top level of the user data Chromium app directory, which is OS-version-dependent. For more information, see logging and user data directory details.

Symbol server

If you are debugging official Google Chrome release builds, use the symbol server:

Tools > Options under Debugging > Symbols. You should set 

 

Alternately, You can set the _NT_SYMBOL_PATH environment variable to include both the Microsoft and Google symbol servers - VS, windbg, and other tools should both respect this environment variable:

_NT_SYMBOL_PATH=SRV*C:symbols*https://msdl.microsoft.com/download/symbols;SRV*C:symbols*https://chromium-browser-symsrv.commondatastorage.googleapis.com

Note that symbol servers will let the debuggers download both the PE files (DLLs and EXEs) and the PDB files.

 

Chrome often loads third party libraries and partial symbols for some of these are also public. For example:

 

AMD: https://download.amd.com/dir/bin

Nvidia: https://driver-symbols.nvidia.com/

Intel: https://software.intel.com/sites/downloads/symbols/

 

For example, for completeness, the following symbol server environment variable will resolve all of the above sources - but this is more than is normally needed:

Source indexing

You should set up source indexing in your debugger (.srcfix in windbg, Tools-> Options-> Debugging-> General-> Enable source server support in Visual Studio) so that the correct source files will automatically be downloaded based on information in the downloaded symbols. This is highly recommended when debugging released Google Chrome builds or looking at crash dumps. Having the correct version of the source files automatically show up saves significant time so you should definitely set this.

Multi-process issues

Chromium can be challenging to debug because of its multi-process architecture. When you select Run in the debugger, only the main browser process will be debugged. The code that actually renders web pages (the Renderer) and the plugins will be in separate processes that's not (yet!) being debugged. ProcessExplorer tool has a process tree view where you can see how these processes are related. You can also get the process IDs associated with each tab from the Chrome Task Manager (right-click on an empty area of the window title bar to open).

Automatically attach to child processes

There are two Visual Studio extensions that enable the debugger to automatically attach to all Chrome processes, so you can debug all of Chrome at once. Microsoft's Child Process Debugging Power Tool is a standalone extension for this, and VsChromium is another option that bundles many other additional features. In addition to installing one of these extensions, you must run Visual Studio as Administrator, or it will silently fail to attach to some of Chrome's child processes.

Single-process mode

One way to debug issues is to run Chromium in single-process mode. This will allow you to see the entire state of the program without extra work (although it will still have many threads). To use single-process mode, add the command-line flag

--single-process

This approach isn't perfect because some problems won't manifest themselves in this mode and some features don't work and worker threads are still spawned into new processes.

Manually attaching to a child process

Tools > Attach and click the  process you want to attach to. Before attaching, make sure you have selected only Native codeSelect...Attach to ProcessNativen operation is not legal in the current state."

When you are debugging multiple processes, open the Debug > Windows > Processes window to switch between them. 

 

 

Sometimes you are debugging something that only happens on startup, and want to see the child process as soon as it starts. Use:

 

 

--renderer-startup-dialog 

You have to disable the sandbox or the dialog box will be prohibited from showing. When the dialog appears, visit Tools > Attach to Process and attach to the process showing the Renderer startup dialog. Now you're debugging in the renderer and can continue execution by pressing OK in the dialog.

 

Startup dialogs also exist for other child process types: --gpu-startup-dialog, --ppapi-startup-dialog, --plugin-startup-dialog (for NPAPI).

 

You can also trythe vs-chromium plug-in to attach to the right processesThe following flags cause child processes to wait for 60 seconds in a busy loop for a debugger to attach to the process. Once either condition is true, it continues on; no exception is thrown.

--wait-for-debugger-children[=filter]

--typerendererpluginppapigpu-process.

When using this option, it may be helpful to limit the number of renderer processes spawned, using:

--renderer-process-limit=1

 

Image File Execution Options

Using Image File Execution Options (IFEO) will not work because CreateProcess() returns the handle to the debugger process instead of the intended child process. There are also issues with the sandbox.

Time travel debugging

You can do time travel debugging using WinDbg Preview (can be installed from the Microsoft Store). This lets you execute a program forward and backwards. This is especially useful when setting data breakpoints (ba command) and reverse continuing, so you can see when a certain variable was last changed to its current value.

JsDbg -- data structure visualization

You can install JsDbg as a plugin for WinDbg or Visual Studio. It interactively lets you look at data structures (such as the DOM tree, Accessibility tree, layout object tree, and others) in a web browser as you debug. See the JsDbg site for some screen shots and usage examples. This also works when examining memory dumps (though not minidumps), and also works together with time travel debugging.

 

Visual Studio hints

//tools/win/DebugVisualizers if you need to modify them (the BUILD.gn file there has additional instructions).

The debugger can be configured to automatically not step into functions bdefault.natstepfilterC:Program Files (x86)Microsoft Visual Studio 14.0Common7PackagesDebuggerVisualizers (for all users)
or %USERPROFILE%My DocumentsVisual Studio 2015Visualizers (for the current user only)

• For Visual Studio 2017 Pro: C:Program Files (x86)Microsoft Visual Studio2017ProfessionalCommon7PackagesDebuggerVisualizers (for all users)
  or %USERPROFILE%My DocumentsVisual Studio 2017Visualizers (for the current user only)

Add regular expressions of functions to not step into. Remember to regex-escape and XML-escape them, e.g. &lt; for < and . for a literal dot. Example:

 

  <Function><Name>operator new</Name><Action>NoStepInto</Action></Function>

  <Function><Name>operator delete</Name><Action>NoStepInto</Action></Function>

  <!-- Skip everything in std -->

  <Function><Name>std::.*</Name><Action>NoStepInto</Action></Function>

  <!-- all methods on WebKit OwnPtr and variants, ... WTF::*Ptr<*>::* -->

  <Function><Name>WTF::.*Ptr&lt;.*&gt;::.*</Name><Action>NoStepInto</Action></Function>

 

This file is read at start of a debugging session (F5), so you don't need to restart Visual Studio after changing it.

More info: Andy Pennel's Blog, microsoft email thread

V8 and Chromium

V8 supports many command-line flags that are useful for debugging. V8 command-line flags can be set via the Chromium command-line flag --js-flags; for instance:

chrome.exe --js-flags="--trace_exception --heap_stats"

Note that some V8 command-line flags exist only in the debug build of V8. For a list of all V8 flags try:

chrome.exe --js-flags="--help"

Graphics debugging

GPU Acceleration of rendering can be more easily debugged with tools. See:

• Graphics Debugging in Visual Studio 2013
• Graphical debugging with NVIDIA NSight

Debugging on another machine

Sometimes it's useful to debug installation and execution on a machine other than your primary build box. To run the installer on said other machine, first build the mini_installer target on your main build machine (e.g., ninja -C outDebug mini_installer). Next, on the debug machine:

• Make the build machine's build volume available on the debug machine either by mounting it locally (e.g., Z:) or by crafting a UNC path to it (e.g., \buildersrc)
• Open up a command prompt and change to a local disk
• Run src oolswincopy-installer.bat in the remote checkout by way of the mount (e.g., Z:PATHTOCHECKOUTsrc...) or UNC path (e.g., \buildersrc...). This will copy the installer, DLLs, and PDBs into your debug machine's C:out or C:uild (depending on if you're rocking the component=shared_library build or not)
• Run C:outDebugmini_installer.exe with the flags of your choice to install Chrome. This can take some time, especially on a slow machine. Watch the Task Manager and wait until mini_installer.exe exits before trying to launch Chrome (by way of the shortcut(s) created by the installer)
• For extra pleasure, add C:outDebug to your _NT_SYMBOL_PATH environment variable

Consider reading the documentation at the top of copy-installer.bat to see how you can run it. It tries to be smart and copy the right things, but you may need to be explicit (e.g., "copy-installer.bat out Debug"). It is safe to re-run the script to copy only modified files (after a rebuild, for example).

 

You can also use the zip action of the isolate scripts (toolsmbmb.py) to package all the files for a target into a single zip file, for example:

 

python toolsmbmb.py zip out/Release base_unittests base_unittests.zip

Finding all memory allocations

It is possible to use Heap Snapshots (requires recent versions of Windows 10?) to get call stacks on all outstanding allocations. This works particularly well if heap snapshots are started as soon as the Chrome browser process is created, but before it starts running. Details can be found in this batch file.

Find memory leaks

The Windows heap manager has a really useful debug flag, where it can be asked to capture and store a stack trace with every allocation. The tool to scrape these stack traces out of processes is UMDH, which comes with WinDbg.

 

UMDH is great. It will capture a snapshot of the heap state as many times as you like, and it'll do it fairly quickly. You then run it again against either a single snapshot, or a pair of snapshots, at which time it'll symbolize the stack traces and aggregate usage up to unique stack traces.

 

Turning on the user stack trace database for chrome.exe with gflags.exe makes it run unbearably slowly; however, turning on the user stack trace database on for the browser alone is just fine.

 

While it's possible to turn on the user stack database with the "!gflag" debugging extension, it's too late to do this by the time the initial debugger breakpoint hits. The only reasonable way to do this is to 

1. Launch GFlags.exe,
2. Enable the user stack trace database (per image below),
3. Launch Chrome under the debugger.
4. Set a breakpont when chrome.dll loads with "sxe ld chrome.dll".
5. Step up, to allow Chrome.dll to initialize.
6. Disable the stack trace database in GFlags.exe.
7. Continue chrome, optionally detaching the debugger.

 

GFlags.exe settings for user mode stack trace database.

If you then ever suffer a browser memory leak, you can snarf a dump of the process with 

 

umdh -p:<my browser pid> > chrome-browser-leak-umdh-dump.txt

 

which can then typically be "trivially" analyzed to find the culprit.

Miscellaneous

Note that until crbug.com/1004989 is fixed you may need to add --disable-features=RendererCodeIntegrity to avoid sandbox crashes in renderer processes when using Application Verifier. See also this page.

• Application Verifier is a free tool from Microsoft (available as part of the Windows SDK) that can be used to flush out programming errors. Starting with M68 Application Verifier can be enabled for chrome.exe without needing to disable the sandbox. After adding chrome.exe to the list of applications to be stressed you need to expand the list of Basics checks and disable the Leak checks. You may also need to disable Handles and Locks checks depending on your graphics driver and specific Chrome version, but the eventual goal is to have Chrome run with Handles and Locks checks enabled. When bugs are found Chrome will trigger a breakpoint so running all Chrome processes under a debugger is recommended. Chrome will run much more slowly because Application Verifier puts every allocation on a separate page.
• You can check the undocumented 'Cuzz' checkbox in Application Verifier to get the Windows thread scheduler to add some extra randomness in order to help expose race conditions in your code.
• Putting every allocation on a separate page will dramatically affect performance so you may want to only do this for some applications. If you right-click on the Heaps checkbox and select Properties you can edit things like the size range for what allocations go into PageHeap (the page-per-allocation system) and you can set a RandRate percentage to randomly put allocations in PageHeap.

• To put a breakpoint on CreateFile(), add this break point:

{,,kernel32.dll}_CreateFileW@28

• {,,kernel32.dll}specifies the DLL (context operator).
• _ prefix means extern "C".
• @28 postfix means _stdcall with the stack pop at the end of the function. i.e. the number of arguments in BYTES.
• You can use DebugView from SysInternals or sawbuck to view LOG() messages that normally goes to stderr on POSIX.