cqueues unusable on aarch64 under LuaJIT

[pspacek]

cqueues on Debian on aarch64 explodes with error:

/usr/bin/luajit -l cqueues -e os.exit(0)

Full build log:
https://buildd.debian.org/status/fetch.php?pkg=knot-resolver&arch=arm64&ver=5.1.2-1&stamp=1596037546&raw=0

This is probably related to some LuaJIT limitations on aarch64 but it effectivelly makes cqueues unusable for applications where performance matters.

[pspacek]

Oh, I forgot to copy&paste the error itself!

/usr/bin/luajit: bad light userdata pointer
stack traceback:
	[C]: at 0xffffb6342ad0
	[C]: in function 'require'
	/usr/share/lua/5.1/cqueues.lua:2: in function </usr/share/lua/5.1/cqueues.lua:1>
	[C]: at 0xaaaae1757d08
	[C]: at 0xaaaae170a4c0

[daurnimator]

Can you confirm that CQS_USE_47BIT_LIGHTUSERDATA_HACK is being set?

Could you provide a C stack trace?

[vcunat]

Debian doesn't provide -dbgsym for this platform (no idea why). When I build it myself from master on the same machine, the error disappears :-/

[daurnimator]

Looks like your debian build could be using an old release:

Get:110 https://deb.debian.org/debian unstable/main arm64 lua-cqueues arm64 20190813-1 [179 kB]

[vcunat]

No, that cqueues version had this fixed already. And my testing was with this anyway:

# apt show lua-cqueues
Package: lua-cqueues
Version: 20200726-1
[...]

[rhenwood-arm]

I'm not close to this project, but have some familiarity with the LuaJIT story on AArch64. MoonJIT has an active AArch64 port - and may be an alternative?

[vcunat]

This is the same with moonjit 2.1.2.

I guess this isn't enough "proof"?

(gdb) bt
#0  lj_err_msg (L=0x18255adee378, em=em@entry=LJ_ERR_BADLU) at lj_err.c:645
#1  0x0000aaaaaab00180 in lua_pushlightuserdata (L=<optimized out>, p=<optimized out>) at lj_state.c:121
#2  0x0000fffff7d29378 in luaopen.cqueues () from /usr/local/lib/lua/5.1/_cqueues.so

[daurnimator]

@vcunat recompile cqueues with -O0; I want to see the exact line in _cqueues.so that is causing the failure.

[vcunat]

As I wrote, I can't reproduce the failure when recompiling by hand (so far). Maybe the address layout is sensitive to something that I don't know about.

[vcunat]

Line isn't surprising: https://salsa.debian.org/lua-team/lua-cqueues/-/blob/2eb74cf0/src/cqueues.c#L2940

[vcunat]

It's the defined(LUA_JITLIBNAME) in the condition. Debian naturally assumes that packages built against vanilla lua 5.1 will be usable with luajit, moonjit, etc.

[vcunat]

For reference, luaossl suffers from the very same (except its version is ancient on Debian anyway).

[daurnimator]

Debian naturally assumes that packages built against vanilla lua 5.1 will be usable with luajit, moonjit, etc.

sadly that isn't true for 64bit platforms. if they want to share the same build for both lua 5.1 and LuaJIT then they'll need to compile with CQS_USE_47BIT_LIGHTUSERDATA_HACK defined.

[pspacek]

@yac Please have a look once you get to Debian packaging.