Re: linux kernel building for mips malta target board

Re: linux kernel building for mips malta target board

[David Daney]

sjhill@realitydiluted.com wrote:
>>>I would like to build latest kernel (ie kernel 2.6.13) for mips malta
>>>board. Can any one advise me the cross-compiler tools version to be 
>>>used for building the linux kernel.
>>>
>>
>>I use both gcc-3.3.1 and gcc-3.4.3 to build 2.6.x mips linux kernels.  I 
>>think for the 2.6.13 kernel any of the latest released versions of 
>>3.3.x, 3.4.x, or 4.0.x would work.  Use the latest binutils (2.16.91 
>>20050817 is what I am using).
>>
> 
> Also, make sure to NOT use any of the 4.1 GCC stuff with Linux/MIPS
> kernels. I am still tracking down errors with it.
> 

Is this the problem you are seeing?:
In file included from include/linux/nfs_fs.h:15,
                  from init/do_mounts.c:12:
include/linux/pagemap.h: In function ‘fault_in_pages_readable’:
include/linux/pagemap.h:236: error: read-only variable ‘__gu_val’ used 
as ‘asm’ output
include/linux/pagemap.h:236: error: read-only variable ‘__gu_val’ used 
as ‘asm’ output
include/linux/pagemap.h:236: error: read-only variable ‘__gu_val’ used 
as ‘asm’ output
include/linux/pagemap.h:236: error: read-only variable ‘__gu_val’ used 
as ‘asm’ output

The compiler behavior has changed since 4.0.1, but I think the new 
behavior is correct.  I am blaming the __get_user macro in 
include/asm-mips/uaccess.h.  It should be possible to fix it there.  The 
alternative is to hack up include/linux/pagemap.h.

David Daney

Re: linux kernel building for mips malta target board

[Stuart Anderson]

On Wed, 2 Nov 2005, David Daney wrote:

> Is this the problem you are seeing?:
> In file included from include/linux/nfs_fs.h:15,
>                 from init/do_mounts.c:12:
> include/linux/pagemap.h: In function \x18fault_in_pages_readable\x19:
> include/linux/pagemap.h:236: error: read-only variable \x18__gu_val\x19 used as 
> \x18asm\x19 output
> include/linux/pagemap.h:236: error: read-only variable \x18__gu_val\x19 used as 
> \x18asm\x19 output
> include/linux/pagemap.h:236: error: read-only variable \x18__gu_val\x19 used as 
> \x18asm\x19 output
> include/linux/pagemap.h:236: error: read-only variable \x18__gu_val\x19 used as 
> \x18asm\x19 output
>
> The compiler behavior has changed since 4.0.1, but I think the new behavior 
> is correct.  I am blaming the __get_user macro in include/asm-mips/uaccess.h. 
> It should be possible to fix it there.  The alternative is to hack up 
> include/linux/pagemap.h.

__get_user() is unhappy, with tpyes that are "const". It uses __typeof()
to create a local variable that it wants to write to. I've intended to
have offer up a patch by now, but, too manyunexpected thing have happened 
in the firs thalf of this week.



                                  Stuart

Stuart R. Anderson                               anderson@netsweng.com
Network & Software Engineering                   http://www.netsweng.com/
1024D/37A79149:                                  0791 D3B8 9A4C 2CDC A31F
                                                   BD03 0A62 E534 37A7 9149

Re: linux kernel building for mips malta target board

[David Daney]

Stuart Anderson wrote:
> On Wed, 2 Nov 2005, David Daney wrote:
> 
>> Is this the problem you are seeing?:
>> In file included from include/linux/nfs_fs.h:15,
>>                 from init/do_mounts.c:12:
>> include/linux/pagemap.h: In function \x18fault_in_pages_readable\x19:
>> include/linux/pagemap.h:236: error: read-only variable \x18__gu_val\x19 
>> used as \x18asm\x19 output
>> include/linux/pagemap.h:236: error: read-only variable \x18__gu_val\x19 
>> used as \x18asm\x19 output
>> include/linux/pagemap.h:236: error: read-only variable \x18__gu_val\x19 
>> used as \x18asm\x19 output
>> include/linux/pagemap.h:236: error: read-only variable \x18__gu_val\x19 
>> used as \x18asm\x19 output
>>
>> The compiler behavior has changed since 4.0.1, but I think the new 
>> behavior is correct.  I am blaming the __get_user macro in 
>> include/asm-mips/uaccess.h. It should be possible to fix it there.  
>> The alternative is to hack up include/linux/pagemap.h.
> 
> 
> __get_user() is unhappy, with tpyes that are "const". It uses __typeof()
> to create a local variable that it wants to write to. I've intended to
> have offer up a patch by now, but, too manyunexpected thing have 
> happened in the firs thalf of this week.
> 

That is my analysis as well.  The typeof converts the const char * into 
a const char which is then unsuitable as the destination in an inline asm.

David Daney

Re: linux kernel building for mips malta target board

[Stuart Anderson]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 1109 bytes --]

On Wed, 2 Nov 2005, David Daney wrote:

>> __get_user() is unhappy, with tpyes that are "const". It uses __typeof()
>> to create a local variable that it wants to write to. I've intended to
>> have offer up a patch by now, but, too many unexpected thing have happened 
>> in the firs thalf of this week.

I shamed myself into sitting down and doing this. 8-)

The attached patch seems to work, or at least doesn't seem to cause
things to blow up. An o32 userspace on a 64-bit kernel comes up
multi-user and can build a kernel, and run a quick subset of LTP.

There was a comment on IRC that there was a register allocation issue which
lead to the current code. I'm not sure of the exact details, but I _think_
this change ends up being equivilent to the code it replaces.



                                 Stuart

Stuart R. Anderson                               anderson@netsweng.com
Network & Software Engineering                   http://www.netsweng.com/
1024D/37A79149:                                  0791 D3B8 9A4C 2CDC A31F
                                                  BD03 0A62 E534 37A7 9149

[-- Attachment #2: uaccess.h.patch --]
[-- Type: TEXT/PLAIN, Size: 1354 bytes --]

Index: uaccess.h
===================================================================
--- uaccess.h	(revision 84)
+++ uaccess.h	(working copy)
@@ -210,17 +210,35 @@
 
 #define __get_user_nocheck(x,ptr,size)					\
 ({									\
-	__typeof(*(ptr)) __gu_val =  (__typeof(*(ptr))) 0;		\
 	long __gu_err = 0;						\
 									\
 	switch (size) {							\
-	case 1: __get_user_asm("lb", ptr); break;			\
-	case 2: __get_user_asm("lh", ptr); break;			\
-	case 4: __get_user_asm("lw", ptr); break;			\
-	case 8: __GET_USER_DW(ptr); break;				\
+	case 1: {							\
+		s8 __gu_val =  (s8) 0;					\
+		__get_user_asm("lb", ptr); 				\
+		(x) = (__typeof__(*(ptr))) __gu_val;			\
+		break;							\
+		}							\
+	case 2:	{							\
+		s16 __gu_val =  (s16) 0;				\
+		__get_user_asm("lh", ptr);				\
+		(x) = (__typeof__(*(ptr))) __gu_val;			\
+		break;							\
+		}							\
+	case 4:	{							\
+		s32 __gu_val = (s32) 0;					\
+		__get_user_asm("lw", ptr);				\
+		(x) = (__typeof__(*(ptr))) __gu_val;			\
+		break;							\
+		}							\
+	case 8:	{							\
+		s64 __gu_val = (s64) 0;					\
+		__GET_USER_DW(ptr);					\
+		(x) = (__typeof__(*(ptr))) __gu_val;			\
+		break;							\
+		}							\
 	default: __get_user_unknown(); break;				\
 	}								\
-	(x) = (__typeof__(*(ptr))) __gu_val;				\
 	__gu_err;							\
 })
 

Re: linux kernel building for mips malta target board

[Ralf Baechle]

> I shamed myself into sitting down and doing this. 8-)
> 
> The attached patch seems to work, or at least doesn't seem to cause
> things to blow up. An o32 userspace on a 64-bit kernel comes up
> multi-user and can build a kernel, and run a quick subset of LTP.
> 
> There was a comment on IRC that there was a register allocation issue which
> lead to the current code. I'm not sure of the exact details, but I _think_
> this change ends up being equivilent to the code it replaces.

It's correct - but triggers plenty of extra warnings and you forgot about
get_user() which has the same kind of issue.  Also you don't have the
guarantee that <linux/types.h> has been included, so in order to avoid a
yet another header file dependency I changed s8, s16 etc. to char, short,
int, long long.  Working on it but as usual uaccess.h is quite a quiz.

  Ralf

Re: linux kernel building for mips malta target board

[Martin Michlmayr]

* Ralf Baechle <ralf@linux-mips.org> [2005-11-03 16:33]:
> > There was a comment on IRC that there was a register allocation issue which
> > lead to the current code. I'm not sure of the exact details, but I _think_
> > this change ends up being equivilent to the code it replaces.
> 
> It's correct - but triggers plenty of extra warnings and you forgot about
> get_user() which has the same kind of issue.  Also you don't have the
> guarantee that <linux/types.h> has been included, so in order to avoid a
> yet another header file dependency I changed s8, s16 etc. to char, short,
> int, long long.  Working on it but as usual uaccess.h is quite a quiz.

What's the status of this?

With current linux-mips git I still get the problem.  As a reminder,
the error is:

  CC      fs/compat_ioctl.o
fs/compat_ioctl.c: In function 'fd_ioctl_trans':
fs/compat_ioctl.c:1831: error: read-only variable '__gu_val' used as 'asm' output
fs/compat_ioctl.c:1831: error: read-only variable '__gu_val' used as 'asm' output
fs/compat_ioctl.c:1831: error: read-only variable '__gu_val' used as 'asm' output
fs/compat_ioctl.c:1831: error: read-only variable '__gu_val' used as 'asm' output
make[1]: *** [fs/compat_ioctl.o] Error 1

-- 
Martin Michlmayr
http://www.cyrius.com/

Re: linux kernel building for mips malta target board

[Ralf Baechle]

On Wed, Jan 18, 2006 at 12:52:03PM +0000, Martin Michlmayr wrote:

> * Ralf Baechle <ralf@linux-mips.org> [2005-11-03 16:33]:
> > > There was a comment on IRC that there was a register allocation issue which
> > > lead to the current code. I'm not sure of the exact details, but I _think_
> > > this change ends up being equivilent to the code it replaces.
> > 
> > It's correct - but triggers plenty of extra warnings and you forgot about
> > get_user() which has the same kind of issue.  Also you don't have the
> > guarantee that <linux/types.h> has been included, so in order to avoid a
> > yet another header file dependency I changed s8, s16 etc. to char, short,
> > int, long long.  Working on it but as usual uaccess.h is quite a quiz.
> 
> What's the status of this?
> 
> With current linux-mips git I still get the problem.  As a reminder,
> the error is:
> 
>   CC      fs/compat_ioctl.o
> fs/compat_ioctl.c: In function 'fd_ioctl_trans':
> fs/compat_ioctl.c:1831: error: read-only variable '__gu_val' used as 'asm' output
> fs/compat_ioctl.c:1831: error: read-only variable '__gu_val' used as 'asm' output
> fs/compat_ioctl.c:1831: error: read-only variable '__gu_val' used as 'asm' output
> fs/compat_ioctl.c:1831: error: read-only variable '__gu_val' used as 'asm' output
> make[1]: *** [fs/compat_ioctl.o] Error 1

I have not been able to find some construct that keeps gcc happy and at the
same time doesn't result in significantly worse code.  That matters because
get_user / put_user are used very often throughout the kernel.

  Ralf

Re: linux kernel building for mips malta target board

[Martin Michlmayr]

* Ralf Baechle <ralf@linux-mips.org> [2006-01-18 13:05]:
> > With current linux-mips git I still get the problem.  As a reminder,
> > the error is:
> > 
> >   CC      fs/compat_ioctl.o
> > fs/compat_ioctl.c: In function 'fd_ioctl_trans':
> > fs/compat_ioctl.c:1831: error: read-only variable '__gu_val' used as 'asm' output
> I have not been able to find some construct that keeps gcc happy and at the
> same time doesn't result in significantly worse code.  That matters because
> get_user / put_user are used very often throughout the kernel.

Who could help with that?  Maciej?  David Daney?

(See the original discussion and the proposed patch at
http://www.spinics.net/lists/mips/msg20671.html)
-- 
Martin Michlmayr
http://www.cyrius.com/