Re: [Minios-devel] [UNIKRAFT RFC PATCH 2/5] plat/kvm: arm64: Enable the fp/simd at the starting point

[Julien Grall <julien@xxxxxxx>]

Hi,

On 21/01/2020 05:13, Jia He wrote:
> Hi Julien
>
> On 2019/12/20 15:50, Julien Grall wrote:
> > On 20/12/2019 01:14, Justin He wrote:
> > > Hi Julien
> > Hello,
> >
> > > > -----Original Message-----
> > > > From: Julien Grall <julien@xxxxxxx>
> > > > Sent: Thursday, December 19, 2019 10:35 PM
> > > > To: Justin He <Justin.He@xxxxxxx>; minios-devel@xxxxxxxxxxxxxxxxxxxx;
> > > > Simon Kuenzer <simon.kuenzer@xxxxxxxxx>; Santiago.Pagani@xxxxxxxxx
> > > > Cc: Felipe Huici <felipe.huici@xxxxxxxxx>; Kaly Xin <Kaly.Xin@xxxxxxx>;
> > > > Julien Grall <Julien.Grall@xxxxxxx>; Sharan.Santhanam@xxxxxxxxx
> > > > Subject: Re: [Minios-devel] [UNIKRAFT RFC PATCH 2/5] plat/kvm: arm64:
> > > > Enable the fp/simd at the starting point
> > > >
> > > > Hi,
> > > >
> > > > On 19/12/2019 14:27, Jia He wrote:
> > > > > Write the sys reg to enable the fp/simd feature, otherwise it will
> > > > > cause floating point/simd exception when touching q0-q31.
> > > > In your first patch you say FPSIMD will be available for application
> > > > only. But here, you will allow Unikraft itself to use it.
> > > >
> > > > So don't you want to turn this only when jumping to the application 
> > > > code?
> > > Yes, this might be more graceful.
> > > But there is a corner case in libukdebug of Unikraft.
> > > If we removed -mgeneral-regs-only, some printf* family functions 
> > > (especially
> > > those have variable length parameters, .e.g va_start/end) will be 
> > > compiled to
> > > use q0-q31.
> > I have already had a lenghty discussion in the past with Wei Chen (see 
> > [1]) about it.
> > If you tell GCC to enable floating point for the kernel, then your 
> > exception path (e.g interrupt) may also use floating. Therefore you 
> > would also have to save/restore them at *every* exception to prevent 
> > any corruption.
> Maybe let me clarify it more clear. Sorry if it caused some 
> misunderstanding.
> In principle, unkraft kernel part can't use fpsimd, but 
> -mgeneral-registers-only can be global
> only. It mean currently I merely can enable/disable 
> -mgeneral-registers-only globally for
> every unikraft source file, even for application part and newlib part.
>
>
> So to resovle it perfectly, seems we need to use different build flag to 
> kernel, application/newlib
> part? E.g.
>
> 1. for unikraft context switch .S file, it shoultn't use 
> -mgeneral-registers-only because we need
> to save/restore v0-v31 registers
>
> 2. for other unikraft kernel source file, we need to use 
> -mgeneral-registers-only to forbid the
> potential abuse.
>
> 3. for application and newlib source file, we need to use 
> -mgeneral-registers-only
I am not sure to follow it. Did you intend to suggest that newlib and 
the application should be compiled with FPSIMD support (i.e. without 
-mgeneral-registers-only)?
> What do you think of it? thanks
I will wait for more clarification before giving more feedback.

Cheers,

--
Julien Grall

_______________________________________________
Minios-devel mailing list
Minios-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/minios-devel