Re: Question about libffi closure

[Andreas Jaeger <aj at suse dot de>]

Andrew Haley <aph@redhat.com> writes:

> Andreas Jaeger writes:
>  > 
>  > On AMD64 the closure tests all fail with a segmentation fauilt.  If
>  > you look e.g. at closure_fn0 you'll see that the trampoline is setup
>  > as:
>  > 
>  > 
>  > The problem is that the trampoline is non-executable, the variables
>  > are declared as static!  If I change the put cl on the stack instead
>  > of static memory (removing the static keyword in line 48 of
>  > closure_fn0.c), the test is successfull and returns the desired
>  > argument:
>  > 
>  > $ LD_LIBRARY_PATH=../.libs/  ./closure_fn0.exe 
>  > 1 2 3 4 127 429 7 8 9 10 11 12 13 19 21 1 3: 680
>  > res: 680
>  > 
>  > My question now is is that test is really correct to allocate the
>  > memory in static memory - and if it is correct, then how can this be
>  > fixed?  Do other platforms have the same problem - and if not, what's
>  > different?
>
>
> It's not a problem, because gcj uses PROT_EXEC.

Glad to hear.

> We can't change to stack allocation in general, because some targets
> don't support executable stack.
>
> The general solution is to use mmap().

So, how shall we fix the testsuite?

Should we do instead of:

  static ffi_closure cl;
  ffi_closure *pcl = &cl;

use:
  ffi_closure *pcl;

  pcl = mmap (NULL, sizeof(ffi_closure), PROT_READ|PROT_EXEC,
              MAP_PRIVATE|MAP_ANONYMOUS, fd, 0);

Ok, I'll try to change the testsuite...

thanks for the hint,
Andreas
-- 
 Andreas Jaeger, aj@suse.de, http://www.suse.de/~aj
  SuSE Linux AG, Deutschherrnstr. 15-19, 90429 Nürnberg, Germany
   GPG fingerprint = 93A3 365E CE47 B889 DF7F  FED1 389A 563C C272 A126

Attachment: pgp00000.pgp
Description: PGP signature