Re: [PATCH 2/3] gdb: improve disassembler styling when Pygments raises an exception

[Andrew Burgess <aburgess@redhat.com>]

Simon Marchi <simark@simark.ca> writes:

>> +# Check that, if the user is using Python Pygments for disassembler
>> +# styling, then the styling correctly switches off when an error is
>> +# detected in the Python code.
>> +proc test_disassembler_error_handling { } {
>> +
>> +    # This test requires the Python Pygments module to be installed
>> +    # and used by GDB.
>> +    if { !$::python_disassembly_styling } {
>> +	return
>> +    }
>> +
>> +    save_vars { env(TERM) } {
>> +	# We need an ANSI-capable terminal to get the output.
>> +	setenv TERM ansi
>> +
>> +	# Restart GDB with the correct TERM variable setting, this
>> +	# means that GDB will enable styling.
>> +	clean_restart_and_disable "restart 4" $::binfile
>> +
>> +	# Disable use of libopcodes for styling.  As this function is
>> +	# only called when Python Pygments module is available, we
>> +	# should now be using that module to style the disassembler
>> +	# output.
>> +	gdb_test_no_output "maint set libopcodes-styling enabled off"
>> +
>> +	# Disassemble a single instruction and ensure that the output
>> +	# has styling markers in it.
>> +	set insn_before [get_single_disassembled_insn]
>> +	gdb_assert { [regexp "\033" $insn_before] } \
>> +	    "have style markers when Pygments is working fine"
>> +
>> +	# Now replace the standard function that colorizes the
>> +	# disassembler output, with a new function that always returns
>> +	# None, this should cause GDB to stop using the Pygments
>> +	# module for disassembler styling.
>> +	gdb_py_test_silent_cmd \
>> +	    [multi_line_input \
>> +		 "python" \
>> +		 "def replacement_colorize_disasm(content,gdbarch):" \
>> +		 "  return None" \
>> +		 "gdb.styling.colorize_disasm = replacement_colorize_disasm" \
>> +		 "\004"] \
>
> Any reason you are using \004 here, instead of end?  I don't quite
> understand why, but it seems to cause some random failures.  Running the
> test under `taskset -c 2` makes it fail most of the time.  Running it
> with check-read1 makes it fail consistently:
>
>   FAIL: gdb.base/style.exp: capture_command_output for x/1i *main
>
> When changing \004 for end, it passes.  I don't have an explanation why
> though.

Turns out the explanation is really simple once you know what it is :)

gdb_py_test_silent_cmd forwards to gdb_test_multiple.  The command I was
sending above, once multi_line_input has done its thing is like this:

  python\ndef ...\n  return None\ngdb.styling....\n\004

This gets sent to gdb_test_multiple, which then sends the command to
GDB, along with a trailing newline - the newline which it thinks will
cause the command to dispatch to GDB.  So our command ends like this:

   ....\n\004\n

Of course, a \004 doesn't need a trailing newline, it is handled
immediately, so that final '\n' is not part of the multi-line python
block.

And so, after \004 has cause the Python block to finish, and the
contents to be processed by the Python interpreter, we still have an
unspent newline character in the input stream, this causes GDB to print
a second prompt, the gdb.log contains:

  (gdb) PASS: gdb.base/style.exp: have style markers when Pygments is working fine
  python
  >def replacement_colorize_disasm(content,gdbarch):
  >  return None
  >gdb.styling.colorize_disasm = replacement_colorize_disasm
  >quit
  (gdb) PASS: gdb.base/style.exp: setup replacement colorize_disasm function
  
  (gdb) FAIL: gdb.base/style.exp: capture_command_output for x/1i *main

And so, we have a race with DejaGnu.  If the two prompts appear slowly
enough, then DejaGnu will match the first one with the end of the 'setup
replacement colorize_disasm function', and the second prompt will be
used to prematurely match with the end of the 'capture_command_output
for x/1i *main' test.

If the two prompts appear quickly then both prompts will be matched as
part of the 'setup replacement colorize_disasm function' test.

The patch I already pushed will resolve this issue completely.

Thanks for finding this bug.

Andrew