Fix exceptions not being inspectable when running binary from PATH

[willhbr]

If the binary is added to $PATH, PROGRAM_NAME is the name of the file,
but not the path to read debug information from. This causes a file open
error, which is incredibly hard to debug. Using Process.executable_path will
instead always be the current executable.

A simple demonstration of this issue is:

def bar
  raise "Oh no it didn't work"
rescue e
  puts e.inspect_with_backtrace
end

bar

Build and move this binary somewhere in your $PATH:

$ crystal build exceptions.cr
$ mv exceptions ~/.local/bin
$ exceptions
Unhandled exception: Error opening file 'exceptions' with mode 'r': No such file or directory (Errno)             
Error while trying to dump the backtrace: Error opening file 'exceptions' with mode 'r': No such file or directory (Errno)

With this patch, the exception is printed as expected:

Oh no it didn't work (Exception)
  from exceptions.cr:2:3 in 'bar'
  from exceptions.cr:11:1 in '__crystal_main'
  from /home/will/projects/gh/crystal-lang/crystal/src/crystal/main.cr:106:5 in 'main_user_code'
  from /home/will/projects/gh/crystal-lang/crystal/src/crystal/main.cr:92:7 in 'main'
  from /home/will/projects/gh/crystal-lang/crystal/src/crystal/main.cr:115:3 in 'main'
  from __libc_start_main
  from _start
  from ???

[willhbr]

Ok I thought alpine was failing because it didn't support /proc/self/exe, but I've changed it to use Process.executable_path and it's still failing. Any advice would be appreciated here!

[rdp]

LGTM, almost wish we'd fix PROGRAM_NAME to always be absolute instead but... :)

Here's what I get on my box without it:

Unhandled exception: Error opening file 'bad4' with mode 'r': No such file or directory (Errno)
Failed to raise an exception: END_OF_STACK
[0x563b1e098ef6] *CallStack::print_backtrace:Int32 +118
[0x563b1e07e786] __crystal_raise +86
[0x563b1e07ecae] ???
[0x563b1e0b6f43] *Crystal::System::File::open<String, String, File::Permissions>:Int32 +163
[0x563b1e0b35e3] *File::new<String, String, File::Permissions, Nil, Nil>:File +67
[0x563b1e0900bd] *CallStack::read_dwarf_sections:(Array(Tuple(UInt64, UInt64, String)) | Nil) +109
[0x563b1e08fe6d] *CallStack::decode_line_number<UInt64>:Tuple(String, Int32, Int32) +45
[0x563b1e08f67e] *CallStack#decode_backtrace:Array(String) +302
[0x563b1e08f532] *CallStack#printable_backtrace:Array(String) +50
[0x563b1e0e5ffd] *Exception+ +77
[0x563b1e0e5e48] *Exception+ +120
[0x563b1e0e3dd9] *AtExitHandlers::run<Int32>:Int32 +441
[0x563b1e0f049b] *Crystal::main<Int32, Pointer(Pointer(UInt8))>:Int32 +139
[0x563b1e089196] main +6
[0x7efeda25ab6b] __libc_start_main +235
[0x563b1e07d6fa] _start +42

If you search for read_dwarf_sections there are a few reports out there :)

[bcardiff]

It’s #8738 do not worry about alpine

[straight-shoota]

I'm not sure this should use PROGRAM_NAME at all. It's a compile time value and not necessarily represents the binary name at runtime.

I'd also like to make this behaviour more resilient in case the file can't be read at all. For example because the executable doesn't exist anymore. We should make sure to still print a meaningful account of the original error.

[ysbaddaden]

I agree with @straight-shoota, and we already do it for Darwin/Mach-O targets. Let's follow up for ELF based targets!

[ysbaddaden]

Suggested change

	Debug::ELF.open(Process.executable_path || PROGRAM_NAME) do |elf|
	return unless path = Process.executable_path
	Debug::ELF.open(path) do |elf|

[willhbr]

Done!

[straight-shoota]

What if exception_path is present but File.open fails?

[willhbr]

Then we're back in the same position, but I think making that more robust should be done in another PR - it'll have to be done for the other executable types as well.

[willhbr]

I've added a check for if the file exists (like the MacOS version), unsure of what to do to indicate that there was a problem reading the backtrace, but I think that can be addressed in another PR.

[RX14]

This is fine, though adding Debug::ELF.open? which returned nil without running the block if the open call failed would be neater.