GDB & LLDB plugins for v8 via debug_helper

[Joyee Cheung]

Hi folks,

I have written a design doc on implementing GDB and LLDB plugins using the debug_helper, so that we can surface more JavaScript states when debugging system core dumps. This would complement the existing Windbg integration that also builds on top of debug_helper.

The initial implementation focuses on annotating JS frames in back traces with JS source information.

The design doc is at https://docs.google.com/document/d/1KkFk_rLbcV7XZ11cnnOolMRixkBqPHLJa7ZiCcjY_t0/edit and the prototype patch is at https://chromium-review.googlesource.com/c/v8/v8/+/7776212 

Feedback welcomed!

Regards,

Joyee

[Olivier Flückiger]

Hi Joyee,

Incidentally I have just started working on fixing various issues and extending --gdbjit. I think a large part of what you are describing is already available behind this flag. As of yesterday it should work again with optimized code and it even shows maglev IR (see --maglev-gdbjit). I am also working on fixing the startup penalty it has, so we might be able to enable it by default on debug builds.

*oli

--
--
v8-dev mailing list
v8-...@googlegroups.com
http://groups.google.com/group/v8-dev
---
You received this message because you are subscribed to the Google Groups "v8-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to v8-dev+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/v8-dev/ECAA6D72-F8BA-446E-9B10-88D5C0E62207%40igalia.com.

[Joyee Cheung]

Hi Olivier,

Thanks, that’s good news to hear. I tried a while back in enabling --gdbjit in Node.js and I think the use cases for the plugins proposed are slightly different from --gdbjit: as far as I know GDB JIT integration has runtime overhead from the additional wiring during code compilation and so needs to be enabled when the application launches. debug_helper on the other hand would also be useful for post-mortem debugging - i.e. debugging crash dumps collected when a release build crashes in a remote environment under usual runtime configuration in V8. The crash dump is essentially just a copy of live memory at the point of the crash and can be configured e.g. at the system level, so it has no runtime overhead and helps navigating sporadic crashes difficult to reproduce locally. Operating on a more complete dump of the process can also make it easier to e.g. investigate memory leaks or corruptions.

Although I think some of the internal machinery maybe shared to make debuggability (no matter live or post-mortem) in general easier to maintain :)

To view this discussion visit https://groups.google.com/d/msgid/v8-dev/CAPfE2j2HYT3PeO8OQ1oe9FGZOSR1NmNaTPFHzZgFi8vL70VKMA%40mail.gmail.com.

[Leszek Swirski]

Hi Joyee,

I think your usecase makes sense, and I think it makes sense for us to take everything that we can get when it comes to improving the debug experience -- I see the work Oli is doing and what you are proposing as complementary, and potentially able to mutually strengthen each other. As I commented on the doc, the only major concern I have with a plugin is making sure that it stays functional when state is corrupted, so that we can always rely on the debugger working when debugging something that is going wrong. Otherwise, big LGTM from my point of view to improve the debugging experience.

- Leszek

To view this discussion visit https://groups.google.com/d/msgid/v8-dev/DDFFDFAD-3A1C-49F0-8B96-E9F4AA26F2D1%40igalia.com.

[Joyee Cheung]

Hi Leszek,

Thanks for the feedback, yes I agree they are complementary - with gdbjit integration there can be more powerful features (e.g. theoretically it may be possible to step through JS-source if sufficient information is provided? That would be quite cool) at the cost of some overhead. debug_helper plugins would provide less information but it helps with cases that can't be reproduced in a setup with gdbjit enabled, and maybe we could implement the coredump-to-heapsnapshot idea that was envisioned in the earlier design of debug_helper. And then there might be cases where GDB has a bug in the paths used by one but not the other, so it’s nice to have a solution that exercises different paths.

Regards,

Joyee

To view this discussion visit https://groups.google.com/d/msgid/v8-dev/CAGxd1t-nQfiDCOZ7KRonZ7k1WntGVJbuQ%2BRVSKD-bN3W_4_cGQ%40mail.gmail.com.