Re: [Xen-devel] RT Xen on ARM - R-Car series

[Julien Grall <julien.grall@xxxxxxx>]

Hi Andrii,

On 2/4/19 3:19 PM, Andrii Anisov wrote:
> On 04.02.19 13:36, Julien Grall wrote:
> > That's a good news! Let me try to address your concerns below one by one.
> Lets do it:)
>
> > > And they employ KPTI enabled kernel in the BSP.
> > KPTI is going to work on Xen. There are no known issue with Linux as 
> > the virtual address is not going to be re-used for other purpose in 
> > the virtual address space. The only inconvenience is the message in 
> > debug build.
> > Just in case, I am not saying it should not be fixed :).
> We have a decision to eliminate it.
>
> > > That reveals another critical issue for us, in addition to Set/Way issue 
> >  From the discussion on the another thread and with other people, this 
> > is not entirely the fault of Xen. This was a misuse of the 
> > instructions by the driver.
> Yep, but it is still on our team to solve.
>
> > While you may want to deal with this in your case, I would like to 
> > avoid promoting bad behavior when using Xen upstream.
> Ah, ok. Will say that in different words.
>
> > > and possible performance drops/irq latency raise due to specter 
> > > mitigation measures.
> > Can you remind me the cores you are using?
> A57+A53.
>
> I see the following in my log:
>
>      (XEN) alternatives: Patching with alt table 00000000002c6608 -> 
> 00000000002c6c80
>      (XEN) CPU0 will call ARM_SMCCC_ARCH_WORKAROUND_1 on exception entry
>      (XEN) CPU2 will call ARM_SMCCC_ARCH_WORKAROUND_1 on exception entry
>      (XEN) CPU3 will call ARM_SMCCC_ARCH_WORKAROUND_1 on exception entry
>      (XEN) CPU1 will call ARM_SMCCC_ARCH_WORKAROUND_1 on exception entry
Cortex-A53 should not be affected by spectre v2, so I imagine they are 
only for A57?
It is going to be hard to disable the workarounds by default. But we can 
consider to provide host-wide or per-guest option to disable them on 
trusted environment.
> > Also, when you mean possible, does it mean you haven't looked the 
> > performance regression?
> We have a preliminary results about performance drop with 
> xen4.12-unstable comparing to a our system with 4.10.
A lot of patches have not been backported in Xen 4.10 (including 
Spectre/Meltdown) that will definitely fix hole but may have an impact 
on the performance. There were not backported because of performance 
reason but because of the complexity of the port and seemly lack of 
interest.
Cheers,

--
Julien Grall

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