parallel bug in 0.35 release

[armgong]

I wrtie rjulia package call julia from R, the following R code run correctly on julia 0.3.3 and 0.4,but fail on 0.35.

library(rjulia)
julia_init()
julia_eval("addprocs(1)")
for (i in 1:2)
{
  julia_void_eval(paste("r=remotecall(",i,", rand, 2, 2)",sep = ""))
  y <- j2r("fetch(r)")
  cat("\n")
  cat("process", i, "got value:\n"); 
  print(y)
}
julia_void_eval("rmprocs(workers())")

on 0.35 show following error

> library(rjulia)
> julia_init()
> julia_eval("addprocs(1)")
NULL
> for (i in 1:2)
+ {
+   julia_void_eval(paste("r=remotecall(",i,", rand, 2, 2)",sep = ""))
+   y <- j2r("fetch(r)")
+   cat("\n")
+   cat("process", i, "got value:\n");
+   print(y)
+ }

process 1 got value:
          [,1]      [,2]
[1,] 0.5080583 0.5226671
[2,] 0.3843354 0.6072951
fatal: error thrown and no exception handler available.
MemoryError()

signal (11): Segmentation fault
unknown function (ip: -1616141115)
unknown function (ip: -1615757548)
unknown function (ip: -1616127477)
unknown function (ip: -1618661289)
unknown function (ip: -1618550229)
unknown function (ip: -1618537003)
unknown function (ip: -1618576727)
unknown function (ip: -1616064465)
unknown function (ip: -1616064242)
unknown function (ip: -1616064028)
unknown function (ip: -1624995090)
unknown function (ip: -1624994754)
jl_trampoline at /data/julia0.3/julia/usr/lib/libjulia.so (unknown line)
jl_apply_generic at /data/julia0.3/julia/usr/lib/libjulia.so (unknown line)
jl_f_apply at /data/julia0.3/julia/usr/lib/libjulia.so (unknown line)
julia_rmprocs_20121 at  (unknown line)
terminate_all_workers at ./multi.jl:1614
jlcall_terminate_all_workers_18781 at /data/julia0.3/julia/usr/bin/../lib/julia/sys.so (unknown line)
jl_apply_generic at /data/julia0.3/julia/usr/lib/libjulia.so (unknown line)
_atexit at ./client.jl:423
jlcall__atexit_18051 at /data/julia0.3/julia/usr/bin/../lib/julia/sys.so (unknown line)
jl_apply_generic at /data/julia0.3/julia/usr/lib/libjulia.so (unknown line)
uv_atexit_hook at /data/julia0.3/julia/usr/lib/libjulia.so (unknown line)
jl_exit at /data/julia0.3/julia/usr/lib/libjulia.so (unknown line)
unknown function (ip: -1624911487)
unknown function (ip: -1624911380)
unknown function (ip: -1624920120)
unknown function (ip: -1547017712)
unknown function (ip: 33247064)

[armgong]

after search code , I think this issue is maybe casued by gc bug ???

[tkelman]

Can you do a git bisect on the release-0.3 branch to determine exactly which commit introduced this?

[armgong]

as @tkelman suggest I try hunt the bug,now found commit 06d01c2 cause this problem.before it everything is ok. with this commit simple code will crash it:

library(rjulia)
julia_init()
julia_void_eval("versioninfo()")

but before this stress test is pass like this:

library(rjulia)
julia_init()
julia_void_eval("versioninfo()")
julia_void_eval("addprocs(2)")
for (j in 1:1000)
{
for (i in 2:3)
{
  julia_void_eval(paste("r=remotecall(",i,", rand, 2, 2)",sep = ""))
  y <- j2r("fetch(r)")
  cat("process",j, i, "got value:\n"); 
  print(y) 
  cat("\n")
}
}
julia_void_eval("rmprocs(workers())")
cat("done\n")

[armgong]

after revert commit 06d01c2 on release-0.3 branch head,everything run like a charm,so julia core dev please revert 06d01c2

[jiahao]

cc @vtjnash

[tkelman]

Very interesting. Thank you for running the bisect @armgong. Such a useful tool, git bisect.

[tkelman]

That particular commit was fixing a different bug, so I'm not sure if just reverting it is the correct solution. It's possible that RJulia might be doing something wrong here with respect to tasks or initialization, but I'll wait for Jameson to weigh in.

[armgong]

if RJulia doing something wrong,it might be in julia_init() ,it only do three thing
1 jl_init(JULIA_HOME)
2 jl_eval_string("Base.init_parallel()")
3 jl_eval_string("Base.init_bind_addr(ARGS)")
because julia 0.3 don't call init parallel function in jlapi.c jl_init_with_image,so we need last two line . 0.4 already have them in jlapi.c jl_init_with_image.

on julia 0.33 add last two line test run ok,if no them can't start worker process.
on julia 0.40 also add last two line (though don't need them since already in jl_init_with_image) test run ok
on julia 0.35 no last two line can't start worker process ,add last two line test run into crash.

[vtjnash]

#9461 will likely fix this when merged

[tkelman]

Is that backportable? It isn't even on master yet, would be good to get it merged sooner rather than later to give it time to be well-tested before backporting into 0.3.6.

[vtjnash]

That pr needs to be rewritten for the new gc, so that's not really an available option to merge to master first