Opteron bug

Opteron bug

[R. J. Wysocki]

Hi,

I hope everyone interested in the development for Opteron is aware of the bug 
described here:

http://www.3dchips.net/content/story.php?id=3927

Yours,
rjw

-- 
Rafael J. Wysocki,
SiSK
[tel. (+48) 605 053 693]
----------------------------
For a successful technology, reality must take precedence over public 
relations, for nature cannot be fooled.
					-- Richard P. Feynman

Re: Opteron bug

[R. J. Wysocki]

On Sunday 20 of June 2004 15:22, Andi Kleen wrote:
> On Sat, 19 Jun 2004 22:29:14 +0200
>
> "R. J. Wysocki" <rjwysocki@sisk.pl> wrote:
> > Hi,
> >
> > I hope everyone interested in the development for Opteron is aware of the
> > bug described here:
> >
> > http://www.3dchips.net/content/story.php?id=3927
>
> The kernel never uses backwards REP prefixes.

So it doesn't matter. :-)

Well, is there any case in which the gcc can produce such stuff?

rjw

Re: Opteron bug

[Jakub Jelinek]

On Sun, Jun 20, 2004 at 01:47:17PM +0200, R. J. Wysocki wrote:
> Well, is there any case in which the gcc can produce such stuff?

GCC doesn't ever generate std instruction (only cld), though users
can use it in inline assembly or assembly source file.
AFAIK x86_64 glibc doesn't use it at all either.

	Jakub

Re: Opteron bug

[Andi Kleen]

On Sat, 19 Jun 2004 22:29:14 +0200
"R. J. Wysocki" <rjwysocki@sisk.pl> wrote:

> Hi,
> 
> I hope everyone interested in the development for Opteron is aware of the bug 
> described here:
> 
> http://www.3dchips.net/content/story.php?id=3927

The kernel never uses backwards REP prefixes.

-Andi

Re: Opteron bug

[Andi Kleen]

On Sun, 20 Jun 2004 13:47:17 +0200
"R. J. Wysocki" <rjwysocki@sisk.pl> wrote:

> On Sunday 20 of June 2004 15:22, Andi Kleen wrote:
> > On Sat, 19 Jun 2004 22:29:14 +0200
> >
> > "R. J. Wysocki" <rjwysocki@sisk.pl> wrote:
> > > Hi,
> > >
> > > I hope everyone interested in the development for Opteron is aware of the
> > > bug described here:
> > >
> > > http://www.3dchips.net/content/story.php?id=3927
> >
> > The kernel never uses backwards REP prefixes.
> 
> So it doesn't matter. :-)

Some user space program could still use it.

You'll have to live with that. CPUs like all other complex
systems have bugs.
 
> Well, is there any case in which the gcc can produce such stuff?

With inline assembly only I guess.

-Andi

Re: Opteron bug

[Jakub Jelinek]

On Mon, Jun 21, 2004 at 02:54:53AM +0300, Denis Vlasenko wrote:
> On Sunday 20 June 2004 15:02, Jakub Jelinek wrote:
> > On Sun, Jun 20, 2004 at 01:47:17PM +0200, R. J. Wysocki wrote:
> > > Well, is there any case in which the gcc can produce such stuff?
> >
> > GCC doesn't ever generate std instruction (only cld), though users
> > can use it in inline assembly or assembly source file.
> > AFAIK x86_64 glibc doesn't use it at all either.
> 
> glibc-2.3/sysdeps/i386/memcopy.h:
> 
> #define BYTE_COPY_BWD(dst_ep, src_ep, nbytes)                                 \
>   do                                                                          \
>     {                                                                         \
>       int __d0;                                                               \
>       asm volatile(/* Set the direction flag, so copying goes backwards.  */  \
>                    "std\n"                                                    \
>                    /* Copy bytes.  */                                         \
>                    "rep\n"                                                    \
>                    "movsb\n"                                                  \
>                    /* Clear the dir flag.  Convention says it should be 0. */ \
>                    "cld" :                                                    \
>                    "=D" (dst_ep), "=S" (src_ep), "=c" (__d0) :                \
>                    "0" (dst_ep - 1), "1" (src_ep - 1), "2" (nbytes) :         \
>                    "memory");                                                 \
>       dst_ep += 1;                                                            \
>       src_ep += 1;                                                            \
>     } while (0)
> 
> WORD_COPY_BWD also does this

I know, but I said x86_64 glibc, which doesn't do this.

	Jakub

Re: [discuss] Re: Opteron bug

[H. Peter Anvin]

Andi Kleen wrote:
> 
> The kernel never uses backwards REP prefixes.
> 

Not even for memmove()?

	-hpa

Re: [discuss] Re: Opteron bug

[Richard B. Johnson]

On Mon, 12 Jul 2004, H. Peter Anvin wrote:

> Andi Kleen wrote:
> >
> > The kernel never uses backwards REP prefixes.
> >
>
> Not even for memmove()?
>
> 	-hpa
>

One can't assume that 'cld' is not required because a user can
execute 'std'. Such a user, calling a kernel function in its
context, could really make a mess by executing 'std' before.

You definitely need a 'cld' inside any copy_to/from_user code.
memmove(), can copy overlapping buffers. It does this by
copying backwards when necessary. If it uses ix86 string
primatives, it executes 'std' for the backwards repeat counts.

Cheers,
Dick Johnson
Penguin : Linux version 2.4.26 on an i686 machine (5570.56 BogoMips).
            Note 96.31% of all statistics are fiction.