Re: [Xen-devel] [PATCH 02/18] xen/arm: Restore HCR_EL2 register

[Marc Zyngier <marc.zyngier@xxxxxxx>]

On 22/03/17 12:45, Mark Rutland wrote:
> On Wed, Mar 22, 2017 at 12:16:20PM +0000, Julien Grall wrote:
>> (CC Mark for the TLB question)
> 
> [Adding Marc since he should understand this better than I do]
> 
> I've trimmed a lot of context here, since it wasn't clear if it was
> relevant to the question. If there's something I've missed, please point
> that out explicitly.
> 
>> I am not entirely sure if we can only restore HCR_RW. My concern is
>> some of the bits are cached in the TLBs. Looking at the ARM ARM
>> (both v7 and v8 see D4.7 in DDI0487A.k_iss10775): "In these
>> descriptions, TLB entries for a translation regime for a particular
>> Exception level are out of context when executing at a higher
>> Exception level.". Which I interpret as TLB result cannot be cached
>> when translating a guest VA to guest PA at EL2. So I guess restoring
>> only HCR_RW might be fine.
> 
> My understanding is that when a translation regime is out-of-context,
> the only requirement is that the core does not begin speculative walks
> for that translation regime. See ARM DDI 0487A.k_iss10775, page D4-1735,
> "Use of out-of-context translation regimes".

+1. An AT instruction is certainly allowed to hit in the TLB, and the
only thing the core is not allowed to do is to speculate (because you
cannot restore a guest context atomically).

> I believe that an explicit address translation involving a translation
> regime can validly make use of (any part of) that regime's state and/or
> allocate new TLB entries for that regime.
> 
> It looks like you're asking if you can avoid installing all of the
> registers related to the EL1&0 regime when performing an EL1&0
> translation at EL2, right?
> 
> I do not believe that is valid; my understanding is that all of the
> relevant registers need to be installed first.

Indeed. I can't see how the you can perform an AT without first
restoring all of the context that define how the page tables are parsed
and including all the element that may be used to hit in the TLBs. Even
worse, you could create TLB entries (as you said, AT doesn't preclude
TLBs from being populated) that would conflict with an existing one on
the next lookup, generating a Conflict Abort.

In short, not restoring the full VM context is doomed.

Now, what is the actual question? ;-)

Cheers,

        M.
-- 
Jazz is not dead. It just smells funny...

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