Re: [PATCH v2 11/40] xen/mpu: build up start-of-day Xen MPU memory region map

[Julien Grall <julien@xxxxxxx>]

On 02/02/2023 10:53, Penny Zheng wrote:
> Hi Julien,
Hi,

> > -----Original Message-----
> > From: Julien Grall <julien@xxxxxxx>
> > Sent: Thursday, February 2, 2023 2:57 AM
> > To: Penny Zheng <Penny.Zheng@xxxxxxx>; xen-devel@xxxxxxxxxxxxxxxxxxxx;
> > Wei Chen <Wei.Chen@xxxxxxx>; Stefano Stabellini
> > <sstabellini@xxxxxxxxxx>; Bertrand Marquis <Bertrand.Marquis@xxxxxxx>;
> > ayan.kumar.halder@xxxxxxxxxx
> > Cc: Volodymyr Babchuk <Volodymyr_Babchuk@xxxxxxxx>
> > Subject: Re: [PATCH v2 11/40] xen/mpu: build up start-of-day Xen MPU
> > memory region map
> >
> > Hi Penny,
> >
> > On 01/02/2023 05:39, Penny Zheng wrote:
> > > > > If we are adding MPU regions in sequence as you suggested, while
> > > > > using bitmap at the same time to record used entry.
> > > > > TBH, this is how I designed at the very beginning internally. We
> > > > > found that if we don't do reorg late-boot to keep fixed in front and
> > > > > switching ones after, each time when we do vcpu context switch, not
> > > > > only we need to hunt down switching ones to disable, while we add
> > > > > new switch-in regions, using bitmap to find free entry is saying
> > > > > that the
> > > > process is unpredictable. Uncertainty is what we want to avoid in
> > > > Armv8-R architecture.
> > > >
> > > > I don't understand why it would be unpredictable. For a given
> > > > combination of platform/device-tree, the bitmap will always look the
> > > > same. So the number of cycles/instructions will always be the same.
> > > In boot-time, it will be always the same. But if we still use bitmap
> > > to find free entry(for switching MPU regions) on runtime, hmmm, I
> > > thought this part will be unpredictable.
> > I know this point is now moot as we agreed on not using a bitmap but I
> > wanted to answer on the unpredictability part.
> >
> > It depends on whether you decide to allocate more entry at runtime. My
> > assumption is you won't and therefore the the time to walk the bitmap will
> > always be consistent.
> In MPU, we don't have something like vttbr_el2 in MMU, to store stage 2
> EL1/EL0 translation table. Xen stage 1 EL2 mapping and stage 2 EL1/EL0
> mapping are both sharing one table.
> So when context switching into different guest, the current design is to disable
> DOM1's guest RAM mapping firstly, then enable DOM2's guest RAM mapping,
> to ensure isolation and safety.
I understood that but I don't understand how this is related to my point 
here. The entries you are replacing are always going to be the same 
after boot.
So if you have a bitmap indicate the fixed entries and you don't add 
more fixed one at runtime, then it will always take the same time to 
walk it.
Cheers,

--
Julien Grall