Re: [PATCH] Fix ppc64 test-fenv

Re: [PATCH] Fix ppc64 test-fenv

[Steve Munroe]

Jakub Jelinek writes

> If the official kernel source contained a buggy prctl so that it is
unsafe
> to try prctl on kernels between 2.4.19 and 2.4.21, then
linux/configure.in
> needs to be changed for 2.4.21 the minimum required kernel version.
> Really, having glibc built without --enable-kernel switch just not work
> forever no matter what kernel it is running on is a bad idea.

prctl is not buggy. It either supports the feature or not. Your patch is OK
and should be integrated.

It just does not correct the fundamental problem. The 2.4.20 kernel does
not
initialize FE0|1 correctly for new processes lwp's. This disables FPE
raise,
until/unless this state is explicitely changed later.

The original fe_nomask.c hack (inherited from PPC32) to set FE0|1 was only
partially reliable on uniprocessors and unreliable on SMPs. That is why we
added the PR_SET_FPEXC feature. I feel that fe_nomask should either work
correctly or return ENOSYS. Your patch prevents test-fenv from seeing a
EINVAL.

The fix (both correct initial state for FE0|1 and the PR_SET_FPEXC feature
are in Marcelo's 2.4.21.pre tree and will be released with 2.4.21.

Re: [PATCH] Fix ppc64 test-fenv

[Roland McGrath]

I've put Jakub's change in.


Thanks,
Roland

Re: [PATCH] Fix ppc64 test-fenv

[Jakub Jelinek]

> Steve Munroe wrote:
> 
> > Actually a bad kernel patch (related to initialing the MSR FE0/1 bits
> > correctly) went in after sometime after the 2.4.19 Suse SLES 8 release and
> > was not fixed again until 2.4.21-pre4 (I think).

If the official kernel source contained a buggy prctl so that it is unsafe
to try prctl on kernels between 2.4.19 and 2.4.21, then linux/configure.in
needs to be changed for 2.4.21 the minimum required kernel version.
Really, having glibc built without --enable-kernel switch just not work
forever no matter what kernel it is running on is a bad idea.

	Jakub

Re: [PATCH] Fix ppc64 test-fenv

[Ulrich Drepper]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Steve Munroe wrote:

> Actually a bad kernel patch (related to initialing the MSR FE0/1 bits
> correctly) went in after sometime after the 2.4.19 Suse SLES 8 release and
> was not fixed again until 2.4.21-pre4 (I think).

That's irrelevant.  What does the official kernel sources contain?

- -- 
- --------------.                        ,-.            444 Castro Street
Ulrich Drepper \    ,-----------------'   \ Mountain View, CA 94041 USA
Red Hat         `--' drepper at redhat.com `---------------------------
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.1 (GNU/Linux)

iD8DBQE+lw5k2ijCOnn/RHQRAp6UAKCvWrqxMb7WtMSzmS3FMo+9LgQS7ACeLnBh
jcoDOGZqyKwlUntibB+FxIU=
=ZUUl
-----END PGP SIGNATURE-----

[PATCH] Fix ppc64 test-fenv

[Jakub Jelinek]

Hi!

Unless ppc64 glibc is built with --enable-kernel=2.4.21 or higher
(without this switch 2.4.19 is assumed), it will fail test-fenv
no matter whatever kernel it is running on.
The following patch fixes it.

2003-04-11  Jakub Jelinek  <jakub@redhat.com>

	* sysdeps/unix/sysv/linux/powerpc/powerpc64/fe_nomask.c
	(__fe_nomask_env): Try prctl even if __ASSUME_NEW_PRCTL_SYSCALL
	is not defined, but the prctl constants are.

--- libc/sysdeps/unix/sysv/linux/powerpc/powerpc64/fe_nomask.c.jj	2003-03-06 12:26:24.000000000 -0500
+++ libc/sysdeps/unix/sysv/linux/powerpc/powerpc64/fe_nomask.c	2003-04-11 10:18:44.000000000 -0400
@@ -27,9 +27,15 @@
 const fenv_t *
 __fe_nomask_env (void)
 {
-#ifdef __ASSUME_NEW_PRCTL_SYSCALL
+#if defined PR_SET_FPEXC && defined PR_FP_EXC_PRECISE
+  int result;
   INTERNAL_SYSCALL_DECL (err);
-  INTERNAL_SYSCALL (prctl, err, 2, PR_SET_FPEXC, PR_FP_EXC_PRECISE);
+  result = INTERNAL_SYSCALL (prctl, err, 2, PR_SET_FPEXC, PR_FP_EXC_PRECISE);
+# ifndef __ASSUME_NEW_PRCTL_SYSCALL
+  if (INTERNAL_SYSCALL_ERROR_P (result, err)
+      && INTERNAL_SYSCALL_ERRNO (result, err) == EINVAL)
+    __set_errno (ENOSYS);
+# endif
 #else  
   __set_errno (ENOSYS);
 #endif

	Jakub