Created attachment 155642 [details]
When hashing strings into tables, Lua 5.2 uses a semi-random
hash seed, which is obtained from a mixture of 'time()' and a
few global addresses (in hope that ASLR would randomize those).
See the 'makeseed' function in 'src/lstate.c' for how it's done.
Since we don't have ASLR, the resulting seed depends only on
time(), which means it can only change once a second, instead
of at every execution.
I think this randomization was meant as a way to prevent DoS
attacks based on hash table collisions (I'm not sure if it's
effective, since the random seed is only computed once at startup).
In any case, the biggest effect this has for me is that performance
of various code parts depends on the ordering of table keys (and
thus, on the seed), but you only get one seed per second, so
running microbenchmarks suddenly becomes a problem.
Now, 'src/lstate.c' provides a way for us to supply a better
seed: we just need to redefine 'luai_makeseed' (by default it's
just a 'time()' call).
The attached patch changes 'luai_makeseed' into an 'arc4random()'
call for both lang/lua52 and lang/lua53 (lua51 doesn't seem to
have an equivalent randomization).
Note that defining __BSD_VISIBLE is only really needed for lua53,
since it defines _XOPEN_SOURCE somewhere above, and without
__BSD_VISIBLE 'arc4random' prototype is not visible with just a
this would have to be validated by upstream first imho
The upstream accepts no patches, certainly not the unportable
ones... Would you like me to to go lua-l and ask if it's OK to
change 'time()' into 'arc4random()' on FreeBSD?
Maybe that's the right thing to do? Let's see what feedback you get.
Any update on this?
Created attachment 164875 [details]
The update is that the upstream doesn't mind .
Also, based on their suggestions, here's a new (less intrusive)
Note that the new patch fixes what seems to be a mistake in
lang/lua53/files/patch-src__Makefile: currently that file changes
"CFLAGS= ..." into "CFLAGS?= ...", which makes it so that the
"..." part is ignored, and the "MYCFLAGS" variable defined from
lang/lua53/Makefile is never used. Note that the equivalent patch
for lua52 (lang/lua53/files/patch-src__Makefile) does not add
the "?", which is the correct way to do things.
Since upstream seems to have no problem with this, are there any other outstanding objections preventing this from being committed?