Skip to content

Latest commit

 

History

History
231 lines (191 loc) · 11.8 KB

Backtracing.rst

File metadata and controls

231 lines (191 loc) · 11.8 KB
Raw

Backtracing support in Swift

When things go wrong, it's always useful to be able to get a backtrace showing where the problem occurred in your program.

Broadly speaking there are three circumstances where you might want a backtrace, namely:

  • Program crashes
  • Runtime errors
  • Specific user-defined program events

Historically, Swift has tended to lean on operating system crash catching support for the first two of these, and hasn't really provided any built-in support for the latter. This is fine for Darwin, where the operating system provides a comprehensive system-wide crash catching facility; it's just about OK on Windows, which also has system-wide crash logging; but it isn't great elsewhere, in particular on Linux where a lot of server-side Swift programs currently rely on a separate package to provide them with some level of backtrace support when errors happen.

What does Swift now support?

Swift now supports:

  • Automatic crash catching and backtrace generation out of the box.
  • Built-in symbolication.
  • A choice of unwind algorithms, including "fast", DWARF and SEH.
  • Interactive(!) crash/runtime error catching.

Crash catching is enabled by default, and won't interfere with any system-wide crash reporters you might be using.

How do I configure backtracing?

There is an environment variable, SWIFT_BACKTRACE, that can be used to configure Swift's crash catching and backtracing support. The variable should contain a ,-separated list of key=value pairs. Supported keys are as follows:

Key Default Meaning
enable yes* Set to no to disable crash catching, or tty to enable only if stdin is a terminal.
demangle yes Set to no to disable demangling.
interactive tty Set to no to disable interaction, or yes to enable always.
color tty Set to yes to enable always, or no to disable. Uses ANSI escape sequences.
timeout 30s Time to wait for interaction when a crash occurs. Setting this to none or 0s will disable interaction.
unwind auto Specifies which unwind algorithm to use. auto means to choose appropriately for the platform. Other options are fast, which does a naïve stack walk; and precise, which uses exception handling data to perform an unwind.
preset auto Specifies which set of preset formatting options to use. Options are friendly, medium or full. auto means to use friendly if interactive, and full otherwise.
sanitize preset If yes, we will try to process paths to remove PII. Exact behaviour is platform dependent.
threads preset Options are all to show backtraces for every thread, or crashed to show only the crashing thread.
registers preset Options are none, all or crashed.
images preset Options are none, all, or mentioned, which only displays images mentioned in a backtrace.
limit 64 Limits the length of the captured backtrace. See below for a discussion of its behaviour. Can be set to none to mean no limit.
top 16 Specify a minimum number of frames to capture from the top of the stack. See below for more.
cache yes Set to no to disable symbol caching. This only has effect on platforms that have a symbol cache that can be controlled by the runtime.
output-to stdout Set to stderr to send the backtrace to the standard error instead of standard output. This may be useful in some CI systems.
swift-backtrace   If specified, gives the full path to the swift-backtrace binary to use for crashes. Otherwise, Swift will locate the binary relative to the runtime library, or using SWIFT_ROOT.

(*) On macOS, this defaults to tty rather than yes.

Backtrace limits

The limit settings are provided both to prevent runaway backtraces and to allow for a sensible backtrace to be produced even when a function has blown the stack through excessive recursion.

Typically in the latter case you want to capture some frames at the top of the stack so that you can see how the recursion was entered, and the frames at the bottom of the stack where the actual fault occurred.

  1. There are limit or fewer frames. In this case we will display all the frames in the backtrace. Note that this _includes_ the case where there are exactly limit frames.
  2. There are more than limit frames.
    1. top is 0. We will display the first limit - 1 frames followed by ... to indicate that more frames exist.
    2. top is less than limit - 1. We will display limit - 1 - top frames from the bottom of the stack, then a ..., then top frames from the top of the stack.
    3. top is greater or equal to limit - 1. We will display ..., followed by limit - 1 frames from the top of the stack.

For example, let's say we have a stack containing 10 frames numbered here 1 to 10, with 10 being the innermost frame. With limit set to 5, you would see:

10
9
8
7
...

With limit set to 5 and top to 2, you would instead see:

10
9
...
2
1

And with limit set to 5 and top to 4 or above, you would see:

...
4
3
2
1

What is the swift-backtrace binary?

swift-backtrace is a program that gets invoked when your program crashes. We do this because when a program crashes, it is potentially in an invalid state and there is very little that is safe for us to do. By executing an external helper program, we ensure that we do not interfere with the way the program was going to crash (so that system-wide crash catchers will still generate the correct information), and we are also able to use any functionality we need to generate a decent backtrace, including symbolication (which might in general require memory allocation, fetching and reading remote files and so on).

You shouldn't try to run swift-backtrace yourself; it has unusual requirements, which vary from platform to platform. Instead, it will be triggered automatically by the runtime.

System specifics

macOS

On macOS, we catch crashes and other events using a signal handler. At time of writing, this is installed for the following signals:

Signal Description Comment
3 SIGQUIT Quit program  
4 SIGILL Illegal instruction  
5 SIGTRAP Trace trap  
6 SIGABRT Abort program  
8 SIGFPE Floating point exception On Intel, integer divide by zero also triggers this.
10 SIGBUS Bus error  
11 SIGSEGV Segmentation violation  

If crash catching is enabled, the signal handler will be installed for any process that links the Swift runtime. If you replace the handlers for any of these signals, your program will no longer produce backtraces for program failures that lead to the handler you have replaced.

Additionally, the runtime will configure an alternate signal handling stack, so that stack overflows can be successfully trapped.

Note that the runtime will not install its signal handlers for a signal if it finds that there is already a handler for that signal. Similarly if something else has already configured an alternate signal stack, it will leave that stack alone.

Once the backtracer has finished handling the crash, it will allow the crashing program to continue and crash normally, which will result in the usual Crash Reporter log file being generated.

Crash catching cannot be enabled for setuid binaries. This is intentional as doing so might create a security hole.

Other Darwin (iOS, tvOS)

Crash catching is not enabled for non-macOS Darwin. You should continue to look at the system-provided crash logs.