[Qemu-devel] checking if linux-user address is mapped (sparc)

[Qemu-devel] checking if linux-user address is mapped (sparc)

[Vince Weaver]

Hello

so now I'm trying to get the vpr spec2k benchmarks going under
sparc32plus-linux-user

In an optimized memory scanning routine the code is doing:

     ldxa  [ %o0 ] #ASI_PNF, %o3

Which is a load from the "no fault" address space.
What this means is that even if you are loading from an invalid address, 
no fault should occur.

qemu of course segfaults when the address is invalid.

Is there some way to check in linux-user if a virtual memory address is 
mapped?  That way it could be checked before the ldl_raw() is run?

hopefully I am not misunderstanding the bug here.

Vince

Re: [Qemu-devel] checking if linux-user address is mapped (sparc)

[Blue Swirl]

[-- Attachment #1: Type: text/plain, Size: 702 bytes --]

On 9/3/08, Vince Weaver <vince@csl.cornell.edu> wrote:
> Hello
>
>  so now I'm trying to get the vpr spec2k benchmarks going under
>  sparc32plus-linux-user
>
>  In an optimized memory scanning routine the code is doing:
>
>     ldxa  [ %o0 ] #ASI_PNF, %o3
>
>  Which is a load from the "no fault" address space.
>  What this means is that even if you are loading from an invalid address, no
> fault should occur.
>
>  qemu of course segfaults when the address is invalid.
>
>  Is there some way to check in linux-user if a virtual memory address is
> mapped?  That way it could be checked before the ldl_raw() is run?
>
>  hopefully I am not misunderstanding the bug here.

Could you test this patch?

[-- Attachment #2: no_fault_accesses.diff --]
[-- Type: plain/text, Size: 2814 bytes --]

Re: [Qemu-devel] checking if linux-user address is mapped (sparc)

[Vince Weaver]

On Wed, 3 Sep 2008, Blue Swirl wrote:
> Could you test this patch?

The patch works great!  vpr works now, and it also fixes the parser 
benchmark.

Thanks for fixing this, I always get completely mixed up when digging 
through the memory access code trying to figure which parts apply to 
linux-user and which parts don't.

Of the 48 Spec2000 benchmarks, we are down to only 2 that don't work 
(facerec and sixtrack).  The remaining problem is floating point related, 
I'll see if I can track it down.

Thanks!

Vince