mips requires alignment for ext2_*_bit functions

mips requires alignment for ext2_*_bit functions

[Dale Farnsworth]

We see immediate alignment failures using reiserfs on mips systems.

The problems go away with the following patch.

-Dale Farnsworth



Index: linux/include/linux/reiserfs_fs.h
===================================================================
RCS file: /cvsdev/hhl-2.4.17/linux/include/linux/reiserfs_fs.h,v
retrieving revision 1.3
diff -u -r1.3 reiserfs_fs.h
--- linux/include/linux/reiserfs_fs.h	2002/01/10 19:46:23	1.3
+++ linux/include/linux/reiserfs_fs.h	2002/03/27 21:59:07
@@ -847,7 +847,7 @@
 #define DEH_Visible 2
 
 /* 64 bit systems (and the S/390) need to be aligned explicitly -jdm */
-#if BITS_PER_LONG == 64 || defined(__s390__) || defined(__hppa__)
+#if BITS_PER_LONG == 64 || defined(__s390__) || defined(__hppa__) || defined(__mips__)
 #   define ADDR_UNALIGNED_BITS  (3)
 #endif
 

Re: mips requires alignment for ext2_*_bit functions

[Oleg Drokin]

Hello!

On Thu, Mar 28, 2002 at 04:43:54PM -0700, Dale Farnsworth wrote:

> We see immediate alignment failures using reiserfs on mips systems.
> The problems go away with the following patch.

Thank you for the report.

The question is should all MIPS processors be threated this way or
is there are smaller subset that suffers from this problems
(and I believe kernel should allow unaligned accesses from inside of the kernel
itself (by emulating perhaps). That's what I learned from lkml recently)

> -#if BITS_PER_LONG == 64 || defined(__s390__) || defined(__hppa__)
> +#if BITS_PER_LONG == 64 || defined(__s390__) || defined(__hppa__) || defined(__mips__)

Bye,
    Oleg

Re: mips requires alignment for ext2_*_bit functions

[Dale Farnsworth]

On Fri, Mar 29, 2002 at 12:20:34PM +0300, Oleg Drokin wrote:
> Hello!
> 
> On Thu, Mar 28, 2002 at 04:43:54PM -0700, Dale Farnsworth wrote:
> 
> > We see immediate alignment failures using reiserfs on mips systems.
> > The problems go away with the following patch.
> 
> Thank you for the report.
> 
> The question is should all MIPS processors be threated this way or
> is there are smaller subset that suffers from this problems
> (and I believe kernel should allow unaligned accesses from inside of the kernel
> itself (by emulating perhaps). That's what I learned from lkml recently)
> 
> > -#if BITS_PER_LONG == 64 || defined(__s390__) || defined(__hppa__)
> > +#if BITS_PER_LONG == 64 || defined(__s390__) || defined(__hppa__) || defined(__mips__)
> 
> Bye,
>     Oleg

Hi Oleg,

Indeed, if unaligned accesses were supported for all accesses across all
archectures, we wouldn't need this #if at all.

Reiserfs uses ext2_set_bit()*.  There is no requirement for this
operation to be atomic.  However, some architectures, including s390 
and little-endian mips, implement it by calling test_and_set_bit(),
which is atomic.  On little-endian mips, test_and_set_bit() uses
the ll and sc instructions which have alignment requirements and
there is no way to efficiently emulate unaligned versions.

* Similarly for ext2_clear_bit(), and ext2_test_bit().

While it may be better in the long run to modify the ext2_*_bit()
functions, adding __mips__ to the current list is expedient and
consistent with the BITS_PER_LONG==64, __s390__, and __hppa__ cases.

-Dale