llnode
llnode copied to clipboard
LLNode Windows support
I thought I’d open this issue to discuss whether Windows support for LLNode is something anyone wants.
I’ve got and LLDB build up and running on Windows so I had a chance to try building LLNode on Windows. Running gyp_llnode worked fairly well.
The resulting Visual Studio project created by gyp required a little tweaking (mostly things I could google for). The main issue was that LLNode uses the getopt library which isn’t on Windows by default. Given that the getopt problem could probably be fixed running LLNode on Windows looks do-able but would need a bit of work and would probably break fairly quickly unless someone was using it regularly.
There’s two reasons why someone might want to do this:
- To debug issues with versions of Node running on Windows.
- To debug issues with core dumps generated on Linux. (LLDB is able to open Linux ELF cores even when it’s running on Mac or Windows.)
I’m not sure how important the first scenario is but the second one might be useful to developers with Windows machines who deploy to Linux. Of course you do have to get LLDB for Windows, probably by building it yourself, which isn’t hard but it is quite time consuming. (Developers with Mac’s who deploy to Linux are effectively already supported.)
Any thoughts? Does anyone else have a strong need for this? (I’m using a Mac so I’m neutral.)
I think this is good addition, but we will probably need to hook it up to node.js CI, so that it will always be tested too. cc @rvagg @Fishrock123
Opened an issue here to discuss the question of how a project might qualify to use our infra: https://github.com/nodejs/build/issues/448, I think it might be a tough call because it'll open floodgates for projects not clearly within the scope of the foundation.
I don't think there is any reason for llnode to not move under Node.js Foundation. If this is a show stopper - what could be done to make this project qualify for this?
I have no problem with llnode to move into nodejs/, perhaps raise an issue in the diagnostics WG and ask if they'd like to take some broad responsibility for it?
@rvagg absolutely!
@hhellyer will this scenario work for you, or would you prefer the project to remain independent?
@rvagg that was "?" not "!", sorry for confusion.
I think the moving the project under Node.js would be great, it would probably improve the visibility and hopefully adoption and participation as well. I'm not sure I know enough about the Node.js build infrastructure for that part of this discussion but I can see why they'd be worried about the load from adding various other projects.
The topic of finding a good home for work form the post-mortem WG is on the agenda for monday. See: https://github.com/nodejs/post-mortem/issues/31. And more specifically this issue is to discuss the issue for work like llnode (as well as others) from the post-mortem WG: https://github.com/nodejs/post-mortem/issues/30
So I was thinking it was more likely in the scope of the post-mortem WG than the diagnostics WG but there is likely some overlap between those WG's so that may not matter.
Trying LLDB/LLNODE on Windows. There seems to be a problem in API/SBDebugger.cpp (LLDB code). It has a hard-coded mangled name for the plugin entry point:
// TODO: mangle this differently for your system - on OSX, the first
// underscore needs to be removed and the second one stays
LLDBCommandPluginInit init_func =
(LLDBCommandPluginInit)dynlib.getAddressOfSymbol(
"_ZN4lldb16PluginInitializeENS_10SBDebuggerE");
Not sure the comment is correct, as we have it working fine on OSX. Using Windows VS2015 the mangled name I get is "?PluginInitialize@lldb@@YA_NVSBDebugger@1@@Z". I'll do some more digging, then maybe raise and fix in LLDB.
Note for reference, re "getopt library isn’t on Windows" above. There is an alternative implementation inside LLDB, see getopt_internal() in lldb/source/Host/common/GetOptInc.cpp.
@rnchamberlain I think if getopt_internal is linked without public exposure we may end up carrying libgetopt with us.
@hhellyer Just curious, were you able to make progress on this?
Apropos getopt, we could vendor the implementation from freebsd or openbsd, it's just one file and it's license-compatible.
I built lldb successfully,but llnode is not support windows now.
I built llnode on windows successfully. @rnchamberlain I use a def file
LIBRARY llnode.dll
EXPORTS
;
; llnode exports
;
_ZN4lldb16PluginInitializeENS_10SBDebuggerE=PluginInitialize
@langhuihui ah, that's clever, so you export a function name from the llnode DLL that matches that hard-coded mangled name in LLDB (which is probably a clang or gcc style mangled name). Easier than fixing it on the LLDB side!
@rnchamberlain @langhuihui great to hear. What would be the next steps in getting this in our testing/validation to make sure we keep it working ?
See https://github.com/nodejs/llnode/issues/137#issuecomment-332460074 - postmortem metadata needs to be enabled in the node binary first.
@rnchamberlain @hhellyer can you help with getting the metadata added to the node binaries. Maybe at least start by adding an issue in the node repo indicating its needed.
Is this all that's needed to add the metadata? https://github.com/cjihrig/node-1/commit/6391209805c79d61fa49bda90a49a211fac49a78
PR to add build support: https://github.com/nodejs/llnode/pull/203 . This allows llnode to be easily built on Windows, though actually debugging will require some more work.
@joaocgreis great to see this come in :)