Re: non-modular 2.6 ppc kernels miscompiled by gcc-3.3.1?

Re: non-modular 2.6 ppc kernels miscompiled by gcc-3.3.1?

[Mikael Pettersson]

On Sun, 12 Oct 2003 21:53:55 +0200, Sam Ravnborg wrote:
>On Sun, Oct 12, 2003 at 09:47:48PM +0200, Mikael Pettersson wrote:
>> Notice __start___ex_table[]'s address: it's not 4-byte aligned.
>> With gcc-3.2.3 it got an 8-byte aligned address in my 2.6 kernel.
>> 
>> vmlinux.lds.S doesn't explicitly align __start___ex_table, so I
>> simply put ". = ALIGN(4);" before it and Voila! now it works.
>
>ld will aling the section according to alingment requirements
>of the symbols inside the section.
>So what happens in your case is that . (current address) is
>un-even. But ld alings the section to a 4-byte boundary,
>due to one of the symbols inside the section.
>
>So the better fix is to define the lables inside the section,
>(read: inside the two '{}').
>
>Care to give my patch a check and report back.

Yes, your patch also works.

/Mikael

Re: non-modular 2.6 ppc kernels miscompiled by gcc-3.3.1?

[Tom Rini]

On Sun, Oct 12, 2003 at 08:19:50PM +0200, Sam Ravnborg wrote:
> 
> On Sun, Oct 12, 2003 at 03:19:22PM +0200, Benjamin Herrenschmidt wrote:
> > Smells like some section alignement issues. Can you check
> > how the __ex_table  section is aligned and where __start___ex_table
> > points to ? (using objdump)
> 
> Or you could try to apply the following patch - it will fix mis-
> alignmnet of above section.

Applied.  Thanks.

-- 
Tom Rini
http://gate.crashing.org/~trini/

Re: non-modular 2.6 ppc kernels miscompiled by gcc-3.3.1?

[Sam Ravnborg]

On Sun, Oct 12, 2003 at 09:47:48PM +0200, Mikael Pettersson wrote:
> Notice __start___ex_table[]'s address: it's not 4-byte aligned.
> With gcc-3.2.3 it got an 8-byte aligned address in my 2.6 kernel.
> 
> vmlinux.lds.S doesn't explicitly align __start___ex_table, so I
> simply put ". = ALIGN(4);" before it and Voila! now it works.

ld will aling the section according to alingment requirements
of the symbols inside the section.
So what happens in your case is that . (current address) is
un-even. But ld alings the section to a 4-byte boundary,
due to one of the symbols inside the section.

So the better fix is to define the lables inside the section,
(read: inside the two '{}').

Care to give my patch a check and report back.

	Sam

Re: non-modular 2.6 ppc kernels miscompiled by gcc-3.3.1?

[Mikael Pettersson]

On Sun, 12 Oct 2003 15:19:22 +0200, Benjamin Herrenschmidt wrote:
>On Sun, 2003-10-12 at 15:10, Mikael Pettersson wrote:
>> When I compile a non-modular 2.6 kernel for ppc with gcc-3.3.1,
>> it oopses in __copy_tofrom_user() [copy_mount_options()] in
>> user-space's first mount /proc call. User-space limps along
>> for a while, oopsing in every mount call, and then hangs.
>> 
>> This occurs with 2.6.0-test5, test6, and the test7-based
>> linuxppc-2.5 tree (rsync:ed today).
>> 
>> Enabling CONFIG_MODULES=y but still keeping everything built-in
>> prevents the oopses.
>
>Smells like some section alignement issues. Can you check
>how the __ex_table  section is aligned and where __start___ex_table
>points to ? (using objdump)

You're right, it turned out to be an alignment issue. Here's what
my System.map looks like with gcc-3.3.1 and CONFIG_MODULES=n:

c017ba4d ? __start___kcrctab
c017ba4d ? __start___kcrctab_gpl
c017ba4d ? __start___ksymtab
c017ba4d ? __start___ksymtab_gpl
c017ba4d ? __stop___kcrctab
c017ba4d ? __stop___kcrctab_gpl
c017ba4d ? __stop___ksymtab
c017ba4d ? __stop___ksymtab_gpl
c017ba4d A __start___ex_table
c017cf08 A __start___bug_table
c017cf08 A __stop___ex_table
c01801c8 A __stop___bug_table

Notice __start___ex_table[]'s address: it's not 4-byte aligned.
With gcc-3.2.3 it got an 8-byte aligned address in my 2.6 kernel.

vmlinux.lds.S doesn't explicitly align __start___ex_table, so I
simply put ". = ALIGN(4);" before it and Voila! now it works.

/Mikael

--- linuxppc-2.5/arch/ppc/kernel/vmlinux.lds.S.~1~	2003-09-28 12:19:36.000000000 +0200
+++ linuxppc-2.5/arch/ppc/kernel/vmlinux.lds.S	2003-10-12 21:16:24.121283928 +0200
@@ -47,6 +47,7 @@
 
   .fixup   : { *(.fixup) }
 
+  . = ALIGN(4);
   __start___ex_table = .;
   __ex_table : { *(__ex_table) }
   __stop___ex_table = .;

Re: non-modular 2.6 ppc kernels miscompiled by gcc-3.3.1?

[Sam Ravnborg]

On Sun, Oct 12, 2003 at 03:19:22PM +0200, Benjamin Herrenschmidt wrote:
> Smells like some section alignement issues. Can you check
> how the __ex_table  section is aligned and where __start___ex_table
> points to ? (using objdump)

Or you could try to apply the following patch - it will fix mis-
alignmnet of above section.

	Sam

===== arch/ppc/kernel/vmlinux.lds.S 1.24 vs edited =====
--- 1.24/arch/ppc/kernel/vmlinux.lds.S	Fri Sep 12 18:26:52 2003
+++ edited/arch/ppc/kernel/vmlinux.lds.S	Sun Oct 12 20:17:25 2003
@@ -47,13 +47,17 @@
 
   .fixup   : { *(.fixup) }
 
-  __start___ex_table = .;
-  __ex_table : { *(__ex_table) }
-  __stop___ex_table = .;
+	__ex_table : {
+		__start___ex_table = .;
+		*(__ex_table)
+		__stop___ex_table = .;
+	}
 
-  __start___bug_table = .;
-  __bug_table : { *(__bug_table) }
-  __stop___bug_table = .;
+	__bug_table : {
+		__start___bug_table = .;
+		*(__bug_table)
+		__stop___bug_table = .;
+	}
 
   /* Read-write section, merged into data segment: */
   . = ALIGN(4096);

Re: non-modular 2.6 ppc kernels miscompiled by gcc-3.3.1?

[Benjamin Herrenschmidt]

On Sun, 2003-10-12 at 15:10, Mikael Pettersson wrote:
> When I compile a non-modular 2.6 kernel for ppc with gcc-3.3.1,
> it oopses in __copy_tofrom_user() [copy_mount_options()] in
> user-space's first mount /proc call. User-space limps along
> for a while, oopsing in every mount call, and then hangs.
> 
> This occurs with 2.6.0-test5, test6, and the test7-based
> linuxppc-2.5 tree (rsync:ed today).
> 
> Enabling CONFIG_MODULES=y but still keeping everything built-in
> prevents the oopses.

Smells like some section alignement issues. Can you check
how the __ex_table  section is aligned and where __start___ex_table
points to ? (using objdump)

Ben.

non-modular 2.6 ppc kernels miscompiled by gcc-3.3.1?

[Mikael Pettersson]

When I compile a non-modular 2.6 kernel for ppc with gcc-3.3.1,
it oopses in __copy_tofrom_user() [copy_mount_options()] in
user-space's first mount /proc call. User-space limps along
for a while, oopsing in every mount call, and then hangs.

This occurs with 2.6.0-test5, test6, and the test7-based
linuxppc-2.5 tree (rsync:ed today).

Enabling CONFIG_MODULES=y but still keeping everything built-in
prevents the oopses.

With gcc-3.2.3 there are no problems.

I don't have a serial console to the ppc box (weird mac serial
port), but I can try to manually type down and decode the first
oops if necessary.

/Mikael