OpenBSD / Sparc

[ts]

With this configuration

obsd% ./ruby -v
ruby 1.9.0 (2004-04-07) [sparc-openbsd3.4]
obsd%

it's normal that ruby use _setjmp()/_longjmp() rather than
setjmp()/longjmp() ?


Guy Decoux

 

[Yukihiro Matsumoto]

Hi,

In message "OpenBSD / Sparc"

| With this configuration
|
|obsd% ./ruby -v
|ruby 1.9.0 (2004-04-07) [sparc-openbsd3.4]
|obsd%
|
| it's normal that ruby use _setjmp()/_longjmp() rather than
| setjmp()/longjmp() ?

Yes, Ruby uses _setjmp()/_longjmp() where available, because we need
not to save signal context. Is there any problem?

matz.

 

[ts]

Y> Yes, Ruby uses _setjmp()/_longjmp() where available, because we need
Y> not to save signal context. Is there any problem?

Just a small problem

obsd% cat b.rb
#!./ruby
require 'timeout'

Timeout::timeout(5) { p 45 }
obsd%

obsd% ~/local/r19/bin/ruby -v b.rb
ruby 1.9.0 (2004-04-07) [sparc-openbsd3.4]
longjmp botch.
Illegal instruction (core dumped)
obsd%

obsd% diff -u eval.c.~1.657.~ eval.c
--- eval.c.~1.657.~ Wed Apr 7 08:30:13 2004
+++ eval.c Thu Apr 8 15:10:02 2004
@@ -111,7 +111,7 @@
#else
typedef jmp_buf rb_jmpbuf_t;
#ifndef setjmp
-#ifdef HAVE__SETJMP
+#ifdef HAVE___SETJMP
#define setjmp(env) _setjmp(env)
#define longjmp(env,val) _longjmp(env,val)
#endif
obsd%

obsd% ./ruby b.rb
45
obsd%

Now on OpenBSD _setjmp() dont save signal context

The setjmp()/longjmp() function pairs save and restore the signal mask
while _setjmp()/_longjmp() function pairs save and restore only the reg-
ister set and the stack (see sigmask(3)).




Guy Decoux

 

[ts]

t> Just a small problem

Well, I can even make work this

obsd% cat c.rb
#!./ruby
require 'net/http'

Net::HTTP.start('moulon.inra.fr', 80) do |http|
r = http.get('/')
p r
end
obsd%

obsd% ./ruby -v c.rb
ruby 1.9.0 (2004-04-07) [sparc-openbsd3.4]
#<Net::HTTPOK 200 OK readbody=true>
obsd%


obsd% ~/local/r19/bin/ruby -v c.rb
ruby 1.9.0 (2004-04-07) [sparc-openbsd3.4]
longjmp botch.
Illegal instruction (core dumped)
obsd%

And with the version of ruby which is in the packages
(http://www.openbsd.org/3.4_packages/sparc.html)
i.e. not compiled by me

obsd% /usr/local/bin/ruby -v c.rb
ruby 1.6.8 (2002-12-24) [sparc-openbsd3.4]
longjmp botch.
Illegal instruction (core dumped)
obsd%


You can't trust "anglois" to compile a program


Guy Decoux

 

[ts]

Y> Yes, Ruby uses _setjmp()/_longjmp() where available, because we need
Y> not to save signal context. Is there any problem?

Well, the problem is perhaps in OpenBSD. One day, perhaps, I'll try with
NetBSD ...


Guy Decoux

 

[YANAGAWA Kazuhisa]

In Message-Id: <[email protected]>

You can't trust "anglois" to compile a program

Well.... just an idea that can be nothing to do with the problem but
did you compiled all version of ruby with same compiler option? My
memory says at least OpenBSD ports can be compiled with stack
protection, and that may cause the problem....

# Again, this is just my thought standing on my short knowledge.

 

[ts]

Y> Well.... just an idea that can be nothing to do with the problem but
Y> did you compiled all version of ruby with same compiler option? My
Y> memory says at least OpenBSD ports can be compiled with stack
Y> protection, and that may cause the problem....

Well, NetBSD and OpenBSD use the same setjmp()/longjmp(), but this is not
true for _setjmp()/_longjmp().

NetBSD use the revision 1.7 with this comment for 1.6

http://cvsweb.netbsd.org/bsdweb.cgi/src/lib/libc/arch/sparc/gen/_setjmp.S

A new implementation of _setjmp/_longjmp that uses the T_FLUSHWIN trap.
This means the stack frame and program counter can restored directly
from the jmp_buf, eliminating the need to backtrack through frames
which does not work in all cases.

OpenBSD use the revision 1.5

http://www.openbsd.org/cgi-bin/cvsweb/src/lib/libc/arch/sparc/gen/_setjmp.S

and when you look at its algorithm, you understand why it can have
problem ...




Guy Decoux