Can I enable debugging mid-simulation?

[joshualmashburn]

Hello! I am working with Simpoint checkpoints I created. I am still new to the standard library, so I am basing my approach off of https://github.com/gem5/gem5/blob/stable/configs/example/gem5_library/checkpoints/simpoints-se-restore.py

I have added a custom debug flag, as I am trying to grab instruction operands and opcodes as they are sent to a certain execution unit (RISCV ISA). I have all that working easily, however, I do not want the warmup instructions included in the trace. The way I envision this is to enable debugging when the script schedules the actual Simpoint interval's instructions, such as on line 144 of simpoints-se-restore.py.

I tried m5.debug.flags['CustomFlag'] = True, but I get the error TypeError: 'AllFlags' object does not support item assignment. Is there a canonical way to do this? I know the --debug-time argument exists, but the checkpoint generated is named after the tick at the start of the warmup, not the start of the Simpoint interval, and I have many of these to do, so I'd like to avoid manual intervention. Thanks!

[joshualmashburn]

I dug around in the m5 source code and found the code where the debug flags are parsed from the command line.
I tried m5.debug.flags['CustomFlag'].enable() and it works! I'm unsure if that's the proper way to do it, but if there's a better way, I'm curious to know.

[powerjg]

This is the right way!

[polpetras]

Hi,

I found this discussion after implementing exactly this feature for our own
simulation workflows. I have a working implementation and would like to
propose it as an upstream contribution.

Design overview:

• gem5 installs a handler on SIGRTMIN+1 at startup (SIGRTMIN is reserved
  for KVM's KVM_KICK_SIGNAL, so +1 avoids the conflict).
• The signal handler follows the existing gem5 async pattern: it sets
  async_event + async_debug_cmd and wakes up the primary event queue. The actual
  command processing runs on the main simulation thread at the next event
  loop iteration — no mid-instruction races.
• Commands are written to /tmp/gem5_debug_<pid>.cmd before sending the
  signal. Supported commands: +FLAG (enable), -FLAG (disable),
  +FLAG:Nt (enable for N ticks then auto-disable), trace:on,
  trace:off.
• A companion shell script util/gem5_debug_ctl.sh handles writing the
  command file and sending the signal, so no C or Python tooling is needed
  on the user side.
• As of now, the feature is Linux-only — real-time signals are not available
  on macOS. The code is guarded with #ifdef SIGRTMIN and compiles cleanly
  on all platforms.
• Flag::_tracing is made std::atomic<bool> (relaxed ordering) to
  eliminate the data race that would otherwise exist when changeFlag() is
  called while other host threads are executing DPRINTF.

I will open a PR shortly. Happy to hear feedback on the design before or
during review.