zxdb: Fuchsia native debugger user guide

This is the command usage guide for zxdb. Please see also:

Quick start

Connecting in-tree

In-tree developers should use the fx debug command to start the debugger. The system must already be running and reachable via networking from your computer:

$ scripts/fx debug
Attempting to start the Debug Agent.
Waiting for the Debug Agent to start.
Connecting (use "disconnect" to cancel)...
Connected successfully.
[zxdb]

The status command will give the current state of the debugger. Be aware if the remote system dies the debugger won't always notice the connection is gone.

Debugging a process or component.

Running a process on Fuchsia is more complicated than in other systems because there are different loader environments (see “A note about launcher environments” below).

The only want to reliably debug all types of processes is to create a filter on the process name via “attach” and start it the normal way you would start that process. The process name is usually the name of the build target that generates it. To check what this is, use “ps” (either in the debugger or from a system shell) with it running.

Note: only the first 32 bytes of the name are included in the Zircon process description. Sometimes the number of path components can cause the name to be truncated. If the filter isn't working, check the actual name in “ps”. We hope to have a better way to match this in the future.

This example sets a pending breakpoint on main to stop at the beginning of execution, and waits for a process called “my_app” to start:

[zxdb] attach my_app
Waiting for process matching "my_app"

[zxdb] break main
Breakpoint 1 (Software) on Global, Enabled, stop=All, @ main
Pending: No matches for location, it will be pending library loads.

Then run the process the way you would in normal use (directly on the command line, via fx test, via the shell's run fuchsia-pkg://..., or another way. The debugger should then immediately break on main (it may take some time to load symbols so you may see a delay before showing the source code):

Attached Process 1 [Running] koid=51590 my_app.cmx
🛑 on bp 1 main(…) • main.cc:222
   220 }
   221
 â–¶ 222 int main(int argc, const char* argv[]) {
   223   foo::CommandLineOptions options;
   224   cmdline::Status status = ParseCommandLine(argc, argv, &options);

You can then do basic commands that are similar to GDB:

next
step
print argv[1]
continue
quit

A note about launcher environments

The following loader environments all have different capabilities (in order from least capable to most capable):

  • The debugger's run <file name> command (base system process stuff).
  • The system console or fx shell (adds some libraries).
  • The base component environment via the shell‘s run and the debugger’s run -c <package url> (adds component capabilities).
  • The test environment via fx test.
  • The user environment when launched from a “story” (adds high-level services like scenic).

This panoply of environments is why the debugger can't have a simple “run” command that always works.

Launching simple command-line processes

Minimal console apps including some unit tests can be launched directly from within the debugger which avoids the “attach” dance:

[zxdb] break main
Breakpoint 1 (Software) on Global, Enabled, stop=All, @ $main
Pending: No matches for location, it will be pending library loads.

[zxdb] run /bin/cowsay

If you get a shared library load error or errors about files or services not being found, it means the app can‘t be run from within the debugger’s launcher environment. This is true even for things that may seem relatively simple.

Directly launching components

Components that can be executed with the console command run fuchsia-pkg://... can be loaded in the debugger with the following command, substituting your component's URL:

[zxdb] run -c fuchsia-pkg://fuchsia.com/your_app#meta/your_app.cmx

Not all components can be launched this way since most higher-level services won‘t be accessible: if you can’t do run ... from the system console, it won‘t work from the debugger either. Note also that fx test is a different environment. According to your test’s dependencies, it may or may not work from the debugger's run command.

Attaching to an existing process

You can attach to most running processes given the process’ KOID. You can get the KOID by running ps on the target Fuchsia system. zxdb also has a built-in ps command:

[zxdb] ps
j: 1030 root
  j: 1079 zircon-drivers
    p: 1926 driver_host:sys
...

Then to attach:

[zxdb] attach 3517
Process 1 Running koid=1249 pwrbtn-monitor

When you’re done, you can choose to detach (keep running) or kill (terminate) the process.

Interaction model

Most command-line debuggers use an exclusive model for input: you’re either interacting with the debugged process’ stdin and stdout, or you’re interacting with the debugger. In contrast, zxdb has an asynchronous model similar to most GUI debuggers. In this model, the user is exclusively interacting with the debugger while arbitrary processes or threads are running or stopped.

When the debugger itself launches a program it will print the program's stdout and stderr to the console. When you attach (either with a filter or with the attach command) they will go to the original place. Currently there is no way to interact with a process’ stdin.

zxdb has a regular noun/verb model for typed commands. The rest of this section gives an overview of the syntax that applies to all commands. Specific commands will be covered in the “Task guide” section below.

Nouns

The possible nouns (and their abbreviations) are:

  • process (pr)
  • job (j)
  • thread (t)
  • frame (f)
  • breakpoint (bp)

Listing nouns

If you type a noun by itself, it lists the available objects of that type:

  • List attached processes

    [zxdb] process
      # State       Koid Name
    â–¶ 1 Not running 3471 debug_agent_unit_tests.cmx
    
  • List attached jobs

    [zxdb] job
      # State   Koid Name
    â–¶ 1 running 3471 sys
    
  • List threads in the current process:

    [zxdb] thread
      # State   Koid Name
    â–¶ 1 Blocked 1348 initial-thread
      2 Blocked 1356 some-other-thread
    
  • List stack frames in the current thread (the thread must be stopped—see pause below):

    [zxdb] frame
    ▶ 0 fxl::CommandLineFromIterators<const char *const *>() • command_line.h:203
      1 fxl::CommandLineFromArgcArgv() • command_line.h:224
      2 main() • main.cc:174
    

Selecting defaults

If you type a noun and its index, you select that as the default for subsequent commands. It also tells you the stats about the new default.

  • Select thread 3 to be the default for future commands:

    [zxdb] thread 3
    Thread 3 Blocked koid=9940 worker-thread
    
  • Select breakpoint 2 to be the default:

    [zxdb] breakpoint 2
    Breakpoint 2 (Software) on Global, Enabled, stop=All, @ MyFunction
    

Verbs

By default, a verb (run, next, print, etc.) applies to the current defaults. So to evaluate an expression in the context of the current stack frame, just type print by itself:

[zxdb] print argv[1]
"--foo=bar"

You can override the default context by prefixing the verb with a noun and its index. So to evaluate an expression in the context of a specific stack frame (in this case, frame 2 of the current thread):

[zxdb] frame 2 print argv[1]
"--foo=bar"

You can keep adding different types of context. This specifies the process, thread, and frame for the print command:

[zxdb] process 1 thread 1 frame 2 print argv[1]
"--foo=bar"

Attributes and settings

Debugger objects have settings associated with them. Use the “get” verb to list the settings for a given object:

[zxdb] breakpoint 1 get
  enabled  true
  location main
  one-shot false
  scope    global
  stop     all
  type     software

The “get” command with a specific attribute will list the attribute and help associated with it:

[zxdb] breakpoint 1 get scope

  ... help text here ...

scope = global

The “set” command sets a value:

[zxdb] breakpoint 1 set scope="process 1 thread 2"
[zxdb] breakpoint 1 set enabled=false

Some settings are hierarchical. A thread inherits settings from its process which in turn inherits settings from the global scope. The “get” command with no context or parameters will list the global settings and the ones for the current process and thread. You can set a global setting to apply to all threads and processes without specific overrides, or override a specific context:

[zxdb] set show-stdout = false            # Applies to all processes with no override.
[zxdb] process 2 set show-stdout = true   # Overrides a specific process.

Some settings are lists. You can use += to append, or specify a new value with “=”. List elements are space-separated (quote strings with spaces).

[zxdb] set symbol-paths = /foo/bar/baz "/home/Dr. Strangelove/cache"
[zxdb] set symbol-paths += /tmp
[zxdb] get symbol-paths
  ... help text ...

symbol-paths =
  • /foo/bar/baz
  • "/home/Dr. Strangelove/cache"
  • /tmp

Attaching and running

Debugging drivers

It's not currently possible to set up the debugger early enough in system startup to debug most driver initialization. And since zxdb itself uses the network, no drivers associated with network communication can be debugged.

Driver debugging support is tracked in issue [5456](<https://bugs.fuchsia.dev/p/fuchsia/issues/detail?id=5456).

You can debug running drivers by attaching like any other process (see “Attaching to an existing process” below). You can delay initialization to allow time to attach by adding a busyloop at the beginning of your code:

volatile bool stop = false;
while (!stop) {}

To break out of the loop after attaching, either set the variable to true:

[zxdb] print stop = true
true
[zxdb] continue

Or jump to the line after the loop:

[zxdb] jump <line #>
[zxdb] continue

Debugging crash dumps

You can load a minidump generated by a crash report. Use the “opendump” verb and supply the local file name of the dump. The debugger must not be attached to another dump or a running system (use “disconnect” first if so).

[zxdb] opendump upload_file_minidump-e71256ba30163a0.dmp
Opening dump file
Dump loaded successfully.

Now the thread, stack, and memory commands can be used to inspect the state of the program. Use “disconnect” to close the dump.

For in-tree users, the fx debug command can take the path to a core file as an argument.

fx debug -c upload_file_minidump-e71256ba30163a0.dmp

Downloading symbols

You can tell zxdb to look for debug symbols for your core dump in a GCS URL and download them automatically. You'll need to run with a few command-line options:

zxdb --symbol-cache $HOME --symbol-server gs://my-bucket-name/namespace

In-tree users will automatically have these options set, with the server pointed to a bucket containing symbols for all release builds.

The first time you use the symbol server, you will have to authenticate using the auth command. The authentication flow will require you to complete part of the authentication in your browser.

[zxdb] auth
To authenticate, please supply an authentication token. You can retrieve a token from:

https://accounts.google.com/o/oauth2/v2/< very long URL omitted >

Once you've retrieved a token, run 'auth <token>'

[zxdb] auth 4/hAF-pASODIFUASDIFUASODIUFSADF329827349872V6
Successfully authenticated with gs://fuchsia-artifacts-release/debug

Debugging multiple processes

You can debug many arbitrary processes at the same time. Attaching or running when a process is already running in the debugger will just create a new one in parallel.

Recall from the “Interaction model” section you can list the current processes with:

[zxdb] process
  # State       Koid Name
â–¶ 1 Running     1249 pwrbtn-monitor
  2 Not running 7235 pwrbtn-monitor

Select one of those as the default by providing its index (not KOID):

[zxdb] process 2

Or apply commands to a specific process (even if it’s not the default) with:

[zxdb] process 2 pause

Running

Working with breakpoints

Breakpoints stop execution when some code is executed. To create a breakpoint, use the break command (b for short) and give it a location:

[zxdb] break main
Breakpoint 3 (Software) on Global, Enabled, stop=All, @ main
   180
 â—‰ 181 int main(int argc, char**argv) {
   182     fbl::unique_fd dirfd;

A location can be expressed in many different ways.

  • Plain function name. This will match functions with the name in any namespace:

    break main
    
  • Member function or functions inside a specific namespace or class:

    break my_namespace::MyClass::MyFunction
    break ::OtherFunction
    
  • Source file + line number (separate with a colon):

    break mymain.cc:22
    
  • Line number within the current frame’s current source file (useful when stepping):

    break 23
    
  • Memory address:

    break 0xf72419a01
    
  • Expression: Prefixing with “*” will treat the following input as an expression that evaluates to an address. This is most often used with hardware breakpoints.

    break --type=write *&foo
    

To list all breakpoints:

[zxdb] breakpoint

Note: this is the “breakpoint” noun (a noun by itself lists the things associated with it). It is not plural.

To clear a specific breakpoint, give that breakpoint index as the context for the clear command (see “Interaction model” above). Here’s we’re using the abbreviation for breakpoint (bp):

[zxdb] bp 2 clear

Or you can clear the current breakpoint:

[zxdb] clear

Whenever you create or stop on a breakpoint, that breakpoint becomes the default automatically so clear always clears the one you just hit. Note that unlike GDB, “clear” takes a breakpoint context before the verb and there are never any arguments after it. Support for GDB-like “clear ” is issue 5452.

Breakpoints can also be enabled or disabled:

[zxdb] disable
[zxdb] bp 4 enable

Other properties can be modified via the “get” and “set” commands (see above for more):

[zxdb] bp 1 set location = Frobulator::GetThing

Hardware data breakpoints (“watchpoints”)

The processor can be set to break execution when it reads or writes certain addresses. This can be particularly useful to track down memory corruption. Create a hardware breakpoint by specifying “write”, “execute” or “read-write” in the “type” for a break command (unlike in some other debuggers, hardware breakpoints are exposed as a type of breakpoint rather than as a separate “watchpoint” concept).

[zxdb] break --type=read-write --size=4 0x12345670

As a shortcut, the “watch” command will take the contents of a variable or the result of an expression and set a data write breakpoint over its range:

[zxdb] watch i
[zxdb] watch foo[5]->bar

Notes:

  • CPUs only support a limited number of hardware watchpoints, typically around 4.

  • The size of a watchpoint range is limited to 1, 2, 4, or 8 bytes and the address must be an even multiple of the size.

  • Unlike GDB, “watch” will evaluate the expression once and set a breakpoint on the result. It won't re-evaluate the expression. In the above example, it will trigger when “bar” changes but not if “foo[5]” changes to point to a different “bar”.

  • If you watch a variable on the stack and nobody touches it, you will often see it hit in another part of the program when the stack memory is re-used. If you get a surprising breakpoint hit, check that execution is still in the frame you expect.

Programmatic breakpoints

You can insert a hardcoded breakpoint in your code if you want to catch some specific condition. Clang has a builtin (it won't work in GCC Zircon builds):

__builtin_debugtrap();

If the debugger is already attached to the process, it will stop as if a normal breakpoint was hit. You can step or continue from there. If the debugger is not already attached, this will cause a crash.

Working with threads

To list the current process’ threads (see “Interaction model” above for more):

[zxdb] thread
  # State   Koid Name
â–¶ 1 Blocked 1323 initial-thread
  2 Running 3462 worker-thread

Often when you attach to a process the thread will be “blocked”, meaning it is stopped on a system call. For asynchronous programs this will typically be some kind of wait.

Most thread control and introspection commands only work when a thread is suspended (not blocked or running). A thread will be suspended when it is stopped at a breakpoint or crashes. You can explicitly suspend a thread with the pause command:

[zxdb] thread 2 pause
🛑 syscalls-x86-64.S:67
   65 m_syscall zx_port_create 60 2 1
   66 m_syscall zx_port_queue 61 2 1
 â–¶ 67 m_syscall zx_port_wait 62 3 0
   68 m_syscall zx_port_cancel 63 3 1
   69 m_syscall zx_timer_create 64 3 1

When a thread is paused the debugger will show the current source code location. Often threads will be in a system call which will resolve to the location in the assembly-language macro file that generated the system call as shown in the above example.

Running pause by itself with no context will pause all threads of all processes currently attached:

[zxdb] pause

Unpause a thread with continue. As before, continue with no context will resume all threads:

[zxdb] continue

Or continue a specific thread:

[zxdb] thread 1 continue

Working with stack frames

A stack frame is a function call. When a function calls another function, a new nested frame is created. So listing the frames of a thread tells you the call stack. You can only see the stack frames when a thread is suspended (see “Working with threads” above).

To list the current thread’s stack frames (the f abbreviation will also work).

[zxdb] frame
▶ 0 fxl::CommandLineFromIterators<const char *const *>() • command_line.h:203
  1 fxl::CommandLineFromArgcArgv() • command_line.h:224
  2 main() • main.cc:174

And to select a given frame as the default:

[zxdb] frame 2

Frames are numbered with “0” being the top of the stack. Increasing numbers go backwards in time.

You can use the up and down commands to navigate the frame list:

[zxdb] up
  1 fxl::CommandLineFromIterators<const char *const *>() • command_line.h:204

[zxdb] down
  0 fxl::CommandLineFromIteratorsFindFirstPositionalArg<const char *const *>() • command_line.h:185

For more context, you can use the backtrace command. This is identical to frame but gives more detailed address information as well as function parameters. This command can be abbreviated bt:

[zxdb] bt
▶ 0 fxl::CommandLineFromIteratorsFindFirstPositionalArg<const char *const *>() • command_line.h:185
      IP = 0x10f982cf2ad0, BP = 0x66b45a01af50, SP = 0x66b45a01af38
      first = (const char* const*) 0x59f4e1268dc0
      last = (const char* const*) 0x59f4e1268dc8
      first_positional_arg = (const char* const**) 0x0
  1 fxl::CommandLineFromIterators<const char *const *>() • command_line.h:204
      IP = 0x10f982cf2ac0, BP = 0x66b45a01af50, SP = 0x66b45a01af40
      first = <'first' is not available at this address. >
      last = <'last' is not available at this address. >
...

Each stack frame has a code location. Use the list command to look at source code. By itself, it lists the source code around the current stack frame’s instruction pointer:

[zxdb] list
   183 inline CommandLine CommandLineFromIteratorsFindFirstPositionalArg(
   184     InputIterator first, InputIterator last,
 â–¶ 185     InputIterator* first_positional_arg) {
   186   if (first_positional_arg)
   187     *first_positional_arg = last;

You can list code around the current instruction pointer of other stack frames, too:

[zxdb] frame 3 list

Or you can list specific things like functions:

[zxdb] list MyClass::MyFunc

File/line numbers:

[zxdb] list foo.cc:43

Or whole files:

[zxdb] list --all myfile.cc:1

Printing values

The print command can evaluate simple C/C++ expressions in the context of a stack frame. When a thread is suspended (see “Working with threads” above) just type:

[zxdb] print i
34

Expressions can use most simple C/C++ syntax:

[zxdb] print &foo->bar[baz]
(const MyStruct*) 0x59f4e1268f70

You can also evaluate expressions in the context of other stack frames without switching to them (see “Interaction model” above for more):

[zxdb] frame 2 print argv[0]
"/bin/cowsay"

Often you will want to see all local variables:

[zxdb] locals
argc = 1
argv = (const char* const*) 0x59999ec02dc0

You can also set variables:

[zxdb] print done_flag = true
true
[zddb] print i = 56
56

Controlling execution (stepping, etc.)

When a thread is suspended (see “Working with threads” above) you can control its execution:

next / n: Advances to the next line, stepping over function calls.

[zxdb] n

step / s: Advances to the next line. If a function call happens before the next line, that function will be stepped into and execution will stop at the beginning of it. You can also supply an argument which is a substring to match of a specific function call. Function names not containing this substring will be skipped and only matching ones will be stepped into:

[zxdb] s
[zxdb] s MyFunction

ss: List function calls on the current line and step in to the call selected, automatically completing any of the other calls that happen to occur first.

[zxdb] ss
  1 std::string::string
  2 MyClass::MyClass
  3 HelperFunctionCall
  4 MyClass::~MyClass
  5 std::string::~string
  quit
>

finish / fi: Exits the function and stops right after the call.

[zxdb] finish

until / u: Given a location (the same as breakpoints, see above), continues the thread until execution gets there. For example, to run until line 45 of the current file:

[zxdb] u 45

jump: Move the instruction pointer to a new address.

[zxdb] jump 22  // Line number
[zxdb] jump 0x87534123  // Address

To run until execution gets back to a given stack frame:

[zxdb] frame 2 until

Assembly language

There are commands that deal with assembly language:

  • disassemble / di: Disassemble at the current location (or a given location)

  • nexti / ni: Step to the next instruction, stepping over function calls.

  • stepi / si: Step the next instruction, following function calls.

  • regs: Get the CPU registers.

zxdb maintains information about whether the last command was an assembly command or a source-code and will show that information on stepping or breakpoint hits. To switch to assembly-language mode, type disassemble, and to switch back to source-code mode, type list.

Low-level memory

  • aspace: Show mapped memory regions.

  • mem-analyze: Dumps memory, trying to interpret pointers.

  • mem-read / x: Dumps memory

  • stack: Provides a low-level analysis of the stack. This is a handy command for low-level debugging.

  • sym-near: Figure out which symbol corresponds to an address.