Meta data for llnode - Node.js version 8

[mhdawson]

In discussions at NodeConf EU I learned that we don't have the meta data needed for llnode in Node.js version 8.

It sounds like the change to the turbofan pipeline in V8 meant that there were so many changes that those who usually maintained the meta data were not able to keep up.

I think we need to

1. Identify who (sadly I don't know off the top of my head who was doing this valuable work recently) had been doing the work in order to get some context on how to do it efficiently.

2. Put together a plan and volunteers to help get the data up to date for Version 8

3. Identify volunteers to keep the data up to date

4. Put testing in place/notifications so that we have CI failures when the data is not up to date. We may want some sort of step that validates this as part of the release process.

@yunong who mentioned this to me
@MylesBorins who asked if we had a in issue covering this.

[joyeecheung]

I am interested in helping this one out. I am working on nodejs/llnode#143 to bring Errors back, and worked on thin strings & stub frames in nodejs/llnode#121 before. Also working on nodejs/llnode#144 to get the CI output the actual errors from lldb instead of silently timing out.

BTW also working on the JS API nodejs/llnode#147 , although need to think about how to compile both/one of the addon and the plugin when installed by npm.

[cjihrig]

Do you have specifics on what isn't working? For example, the errors implementation seems to have changed in V8 like Joyee mentioned, but as of nodejs/llnode@3ab289b, all of the postmortem metadata appeared to be loading properly, and the test suite was passing with TurboFan (on my machine anyway - we do need a better CI story there).

[mhdawson]

@yunong can you list what you was not working for you ?

[mhdawson]

ping @yunong

[mhdawson]

@cjihrig do you know who has been updating the metadata (since it sounds like it has been done). Would like to understand the process and see if we have it documented somewhere.

[cjihrig]

I did the bulk of the metadata update. As far as I know, it's not documented anywhere. It mostly involves seeing what metadata no longer exists (which llnode will tell you) and then jumping through V8 commits to find out what happened to it.

A big problem is that the metadata can disappear from V8 and Node, and doesn't present any issues until someone is trying to do postmortem debugging with llnode. I'll work on putting together a test in core to try to catch this scenario. Another option is to make compilation fail when certain constants don't exist. This currently happens on SmartOS - constants disappear, and src/v8abbr.h or the ustack helper simply don't compile.

[mhdawson]

@cjihrig thanks, it would be a good step to have tests that will tell when we have problem with the metadata.

How often is has an update needed ? If it is for most V8 versions then we probably want to try to have the update process integrated into the steps used to update V8 in core.

[bnoordhuis]

Practically every major V8 upgrade and infrequently after minor or patch updates.

Paradoxically, before the V8 team started maintaining the postmortem script we knew about most breaking changes because the build would break after an upgrade. :-)

to have the update process integrated into the steps used to update V8 in core

llnode is complex enough that I wouldn't want to inflict that on someone upgrading V8.

[jclulow]

to have the update process integrated into the steps used to update V8 in core

llnode is complex enough that I wouldn't want to inflict that on someone upgrading V8.

One might argue that by updating V8 without doing the work to ensure the post mortem metadata is correct, you're actually inflicting broken software changes on users.

Ensuring the post mortem bits still function is part of the real cost of updating V8, and the process for updating V8 should reflect that.

[bnoordhuis]

Easy to say when you're not doing the work. If Joyent has an interest in postmortem work, they should at least try a little harder to keep the upstream smartos build in working shape.

[jclulow]

This isn't really a Joyent-specific issue at all. We tend to use mdb_v8; this issue seems to be centred on llnode, a similar but unrelated body of software that also consumes the post mortem metadata.

[bnoordhuis]

I was specifically referring to nodejs/node-v8#21, which is smartos-specific and was ultimately fixed by people with zero interest in that platform.

I'll make you a deal: your company makes upgrades frictionless on the smartos front and we'll look into keeping the postmortem metadata in sync across upgrades. That's in your best interest too, right?

[jclulow]

I'm not really sure what motivates your apparent hostility, but nonetheless would like to note that @cjihrig works for Joyent. You'll see him earlier in this thread, where he talks about working on the metadata update. He also seems to have engaged on, and fixed, the other issue you're referencing.