Xen devel - Sep 2012 - [PATCH mini-os enhancements for vtpm 7/8] add floating point and sse to mini-os

This patch adds floating point and sse support to mini-os by
initializing the floating point unit and the see unit during domain boot up.


Signed of by: Matthew Fioravante matthew.fioravante@jhuapl.edu

diff --git a/extras/mini-os/arch/x86/setup.c
b/extras/mini-os/arch/x86/setup.c
--- a/extras/mini-os/arch/x86/setup.c
+++ b/extras/mini-os/arch/x86/setup.c
@@ -74,9 +74,28 @@ shared_info_t *map_shared_info(unsigned long pa)
     return (shared_info_t *)shared_info;
 }
 
+static inline void fpu_init(void) {
+    asm volatile("fninit");
+}
+
+#ifdef __SSE__
+static inline void sse_init(void) {
+    unsigned long status = 0x1f80;
+    asm volatile("ldmxcsr %0" : : "m" (status));
+}
+#else
+#define sse_init()
+#endif
+
 void
 arch_init(start_info_t *si)
 {
+    /*Initialize floating point unit */
+        fpu_init();
+
+        /* Initialize SSE */
+        sse_init();
+
     /* Copy the start_info struct to a globally-accessible area. */
     /* WARN: don''t do printk before here, it uses information from
        shared_info. Use xprintk instead. */
@@ -99,6 +118,7 @@ arch_init(start_info_t *si)
         (unsigned long)failsafe_callback, 0);
 #endif
 
+
 }
 
 void




_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel

Matthew Fioravante, le Mon 17 Sep 2012 18:04:47 -0400, a écrit
:
> This patch adds floating point and sse support to mini-os by
> initializing the floating point unit and the see unit during domain boot
up.
> 
> 
> Signed of by: Matthew Fioravante matthew.fioravante@jhuapl.edu
I don''t remember: does xen support optimizing out FPU context switch
when a domain does not use the FPU?  It might be useful to make FPU
initialization optional in that case, to keep context switch faster.
Apart from that,

Acked-by: Samuel Thibault <samuel.thibault@ens-lyon.org>

At 00:20 +0200 on 18 Sep (1347927637), Samuel Thibault
wrote:
> Matthew Fioravante, le Mon 17 Sep 2012 18:04:47 -0400, a écrit :
> > This patch adds floating point and sse support to mini-os by
> > initializing the floating point unit and the see unit during domain
boot up.
> > 
> > 
> > Signed of by: Matthew Fioravante matthew.fioravante@jhuapl.edu
> 
> I don''t remember: does xen support optimizing out FPU context
switch
> when a domain does not use the FPU? 
Yes, but using the FPU once at boot shouldn''t interfere with that.

Tim.

Tim Deegan, le Tue 18 Sep 2012 01:03:22 +0100, a écrit :
> At 00:20 +0200 on 18 Sep (1347927637), Samuel Thibault wrote:
> > Matthew Fioravante, le Mon 17 Sep 2012 18:04:47 -0400, a écrit :
> > > This patch adds floating point and sse support to mini-os by
> > > initializing the floating point unit and the see unit during
domain boot up.
> > > 
> > > 
> > > Signed of by: Matthew Fioravante matthew.fioravante@jhuapl.edu
> > 
> > I don''t remember: does xen support optimizing out FPU context
switch
> > when a domain does not use the FPU? 
> 
> Yes, but using the FPU once at boot shouldn''t interfere with that.
Ah, right, the initialize context isn''t restored until FPU is really
used.  Then it''s all OK.

Samuel

On Mon, 2012-09-17 at 23:04 +0100, Matthew Fioravante
wrote:
> This patch adds floating point and sse support to mini-os by
> initializing the floating point unit and the see unit during domain boot
up.
> 
> 
> Signed of by: Matthew Fioravante matthew.fioravante@jhuapl.edu
> 
> diff --git a/extras/mini-os/arch/x86/setup.c
> b/extras/mini-os/arch/x86/setup.c
> --- a/extras/mini-os/arch/x86/setup.c
> +++ b/extras/mini-os/arch/x86/setup.c
> @@ -74,9 +74,28 @@ shared_info_t *map_shared_info(unsigned long pa)
>      return (shared_info_t *)shared_info;
>  }
>  
> +static inline void fpu_init(void) {
> +    asm volatile("fninit");
> +}
> +
> +#ifdef __SSE__
How and when is this symbol defined?
> +static inline void sse_init(void) {
> +    unsigned long status = 0x1f80;
> +    asm volatile("ldmxcsr %0" : : "m" (status));
> +}
> +#else
> +#define sse_init()
> +#endif
> +
>  void
>  arch_init(start_info_t *si)
>  {
> +    /*Initialize floating point unit */
> +        fpu_init();
> +
> +        /* Initialize SSE */
> +        sse_init();
The indentation here is a bit suspect.
> +
>      /* Copy the start_info struct to a globally-accessible area. */
>      /* WARN: don''t do printk before here, it uses information
from
>         shared_info. Use xprintk instead. */
> @@ -99,6 +118,7 @@ arch_init(start_info_t *si)
>          (unsigned long)failsafe_callback, 0);
>  #endif
>  
> +
>  }
>  
>  void
> 
>

Ian Campbell, le Tue 18 Sep 2012 08:21:18 +0100, a écrit
:
> On Mon, 2012-09-17 at 23:04 +0100, Matthew Fioravante wrote:
> > This patch adds floating point and sse support to mini-os by
> > initializing the floating point unit and the see unit during domain
boot up.
> > 
> > 
> > Signed of by: Matthew Fioravante matthew.fioravante@jhuapl.edu
> > 
> > diff --git a/extras/mini-os/arch/x86/setup.c
> > b/extras/mini-os/arch/x86/setup.c
> > --- a/extras/mini-os/arch/x86/setup.c
> > +++ b/extras/mini-os/arch/x86/setup.c
> > @@ -74,9 +74,28 @@ shared_info_t *map_shared_info(unsigned long pa)
> >      return (shared_info_t *)shared_info;
> >  }
> >  
> > +static inline void fpu_init(void) {
> > +    asm volatile("fninit");
> > +}
> > +
> > +#ifdef __SSE__
> 
> How and when is this symbol defined?
This is defined by the compiler if some -msse is enabled. This is the
default on i686/x86_64 at least.

Samuel

On Tue, 2012-09-18 at 08:31 +0100, Samuel Thibault
wrote:
> Ian Campbell, le Tue 18 Sep 2012 08:21:18 +0100, a écrit :
> > On Mon, 2012-09-17 at 23:04 +0100, Matthew Fioravante wrote:
> > > This patch adds floating point and sse support to mini-os by
> > > initializing the floating point unit and the see unit during
domain boot up.
> > > 
> > > 
> > > Signed of by: Matthew Fioravante matthew.fioravante@jhuapl.edu
> > > 
> > > diff --git a/extras/mini-os/arch/x86/setup.c
> > > b/extras/mini-os/arch/x86/setup.c
> > > --- a/extras/mini-os/arch/x86/setup.c
> > > +++ b/extras/mini-os/arch/x86/setup.c
> > > @@ -74,9 +74,28 @@ shared_info_t *map_shared_info(unsigned long
pa)
> > >      return (shared_info_t *)shared_info;
> > >  }
> > >  
> > > +static inline void fpu_init(void) {
> > > +    asm volatile("fninit");
> > > +}
> > > +
> > > +#ifdef __SSE__
> > 
> > How and when is this symbol defined?
> 
> This is defined by the compiler if some -msse is enabled. This is the
> default on i686/x86_64 at least.
Do we support processors/compilers without SSE or conversely do we
require SSE support?

It sounds like either we need a runtime test or the compile time test is
unnecessary.

Ian.

Ian Campbell, le Tue 18 Sep 2012 08:38:23 +0100, a écrit
:
> On Tue, 2012-09-18 at 08:31 +0100, Samuel Thibault wrote:
> > Ian Campbell, le Tue 18 Sep 2012 08:21:18 +0100, a écrit :
> > > On Mon, 2012-09-17 at 23:04 +0100, Matthew Fioravante wrote:
> > > > This patch adds floating point and sse support to mini-os by
> > > > initializing the floating point unit and the see unit during
domain boot up.
> > > > 
> > > > 
> > > > Signed of by: Matthew Fioravante
matthew.fioravante@jhuapl.edu
> > > > 
> > > > diff --git a/extras/mini-os/arch/x86/setup.c
> > > > b/extras/mini-os/arch/x86/setup.c
> > > > --- a/extras/mini-os/arch/x86/setup.c
> > > > +++ b/extras/mini-os/arch/x86/setup.c
> > > > @@ -74,9 +74,28 @@ shared_info_t *map_shared_info(unsigned
long pa)
> > > >      return (shared_info_t *)shared_info;
> > > >  }
> > > >  
> > > > +static inline void fpu_init(void) {
> > > > +    asm volatile("fninit");
> > > > +}
> > > > +
> > > > +#ifdef __SSE__
> > > 
> > > How and when is this symbol defined?
> > 
> > This is defined by the compiler if some -msse is enabled. This is the
> > default on i686/x86_64 at least.
> 
> Do we support processors/compilers without SSE or conversely do we
> require SSE support?
> 
> It sounds like either we need a runtime test or the compile time test is
> unnecessary.
If the code is compiled with -msse (and thus __SSE__ defined), the
compiler itself can emit sse code, so a runtime test will not be enough
to avoid a crash on a non-sse machine.

The compile test is however not strictly needed since the code is
provided in assembly. And it might be useful to enable sse anyway in
case some -msse code gets pulled. IIRC we only support SSE machines, but
I''m not 100% sure.

Samuel