Re: [Tarantool-patches] [PATCH luajit 0/2] Trace abort on FFI sandwich or mode change

[Igor Munkin <imun@tarantool.org>]

Vlad,

Thanks for the review!

On 03.04.20, Vladislav Shpilevoy wrote:
> Hi! Thanks for the patch!
> 
> I am getting build errors here:
> 
> Undefined symbols for architecture x86_64:
>   "__Unwind_DeleteException", referenced from:
>       _lj_err_unwind_dwarf in libluajit.a(lj_err.o)
>   "__Unwind_GetCFA", referenced from:
>       _lj_err_unwind_dwarf in libluajit.a(lj_err.o)
>   "__Unwind_RaiseException", referenced from:
>       _lj_err_throw in libluajit.a(lj_err.o)
>   "__Unwind_SetGR", referenced from:
>       _lj_err_unwind_dwarf in libluajit.a(lj_err.o)
>   "__Unwind_SetIP", referenced from:
>       _lj_err_unwind_dwarf in libluajit.a(lj_err.o)
> 
> I think they are not related to your patch, but
> probably you know how to fix them? Did I miss a build
> option? I used 'make -j'.

Hm, it totally doesn't relate to the patch. AFAIR you use MacOS, so the
problems might be related to the external unwinder ([1], sec. 6.2) used
by luajit. Could you please share your build environment configuration?

> 
> I saw you fixed some things requested by Sergey, but I don't
> see them on the branch. Did you push the latest version? For
> example, in the first commit you had a typo in the commit
> message - 'whe' instead of 'when'. I still see 'whe', on github
> too.

Sorry, my fault. I had been waiting for Sergos comments related to my
post-review fixes, avoiding several force pushes. Our discussion had
been stalled for a while, and I've just received his LGTM and I've
pushed the fixed patches to the remote branch. Next time I'll update the
upstream ASAP.

Please, let me know, whether I need to send the actual series version.

> 
> On 27/03/2020 11:47, Igor Munkin wrote:
> > This series closes two issues related to the JIT machinery behaviour:
> > * "FFI sandwich"(*) detection is introduced. If sandwich is detected
> >   while trace recording the recording is aborted. The sandwich detected
> >   while mcode execution leads to the platform panic.
> > * luaJIT_setmode call is prohibited while mcode execution and leads to
> >   the platform panic.
> > 
> > (*) The following stack mix is called FFI sandwich.
> >     | Lua-FFI -> С routine -> Lua-C API -> Lua VM
> >     This sort of re-entrancy is explicitly not supported by LuaJIT
> >     compiler. For more info see [1].
> > 
> > Branch: https://github.com/tarantool/luajit/tree/imun/ffi-sandwich
> > 
> > [1]: https://github.com/tarantool/tarantool/issues/4427
> > 
> > Igor Munkin (2):
> >   jit: abort trace recording and execution for C API
> >   jit: abort trace execution on JIT mode change
> > 
> >  src/lj_api.c                             | 35 ++++++++++----
> >  src/lj_dispatch.c                        |  5 ++
> >  src/lj_errmsg.h                          |  2 +
> >  test/gh-4427-ffi-sandwich/CMakeLists.txt |  1 +
> >  test/gh-4427-ffi-sandwich/libsandwich.c  | 59 ++++++++++++++++++++++++
> >  test/gh-4427-ffi-sandwich/test.lua       | 26 +++++++++++
> >  test/lj-flush-on-trace/CMakeLists.txt    |  1 +
> >  test/lj-flush-on-trace/libflush.c        | 31 +++++++++++++
> >  test/lj-flush-on-trace/test.lua          | 25 ++++++++++
> >  9 files changed, 176 insertions(+), 9 deletions(-)
> >  create mode 100644 test/gh-4427-ffi-sandwich/CMakeLists.txt
> >  create mode 100644 test/gh-4427-ffi-sandwich/libsandwich.c
> >  create mode 100644 test/gh-4427-ffi-sandwich/test.lua
> >  create mode 100644 test/lj-flush-on-trace/CMakeLists.txt
> >  create mode 100644 test/lj-flush-on-trace/libflush.c
> >  create mode 100644 test/lj-flush-on-trace/test.lua
> > 

[1]: https://www.uclibc.org/docs/psABI-x86_64.pdf

-- 
Best regards,
IM