Re: [PATCH v5 3/3] ppc: Enable 2nd DAWR support on p10

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 3978 bytes --]

On Wed, Apr 21, 2021 at 12:24:22PM +0530, Ravi Bangoria wrote:
> Hi Cedric,
> 
> On 4/21/21 12:01 PM, Cédric Le Goater wrote:
> > On 4/21/21 8:20 AM, Ravi Bangoria wrote:
> > > Hi David,
> > > 
> > > On 4/19/21 10:23 AM, David Gibson wrote:
> > > > On Mon, Apr 12, 2021 at 05:14:33PM +0530, Ravi Bangoria wrote:
> > > > > As per the PAPR, bit 0 of byte 64 in pa-features property indicates
> > > > > availability of 2nd DAWR registers. i.e. If this bit is set, 2nd
> > > > > DAWR is present, otherwise not. Use KVM_CAP_PPC_DAWR1 capability to
> > > > > find whether kvm supports 2nd DAWR or not. If it's supported, allow
> > > > > user to set the pa-feature bit in guest DT using cap-dawr1 machine
> > > > > capability. Though, watchpoint on powerpc TCG guest is not supported
> > > > > and thus 2nd DAWR is not enabled for TCG mode.
> > > > > 
> > > > > Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
> > > > > Reviewed-by: Greg Kurz <groug@kaod.org>
> > > > 
> > > > So, I'm actually not sure if using an spapr capability is what we want
> > > > to do here.  The problem is that presumably the idea is to at some
> > > > point make the DAWR1 capability default to on (on POWER10, at least).
> > > > But at that point you'll no longer to be able to start TCG guests
> > > > without explicitly disabling it.  That's technically correct, since we
> > > > don't implement DAWR1 in TCG, but then we also don't implement DAWR0
> > > > and we let that slide... which I think is probably going to cause less
> > > > irritation on balance.
> > > 
> > > Ok. Probably something like this is what you want?
> > > 
> > > Power10 behavior:
> > >    - KVM does not support DAWR1: Boot the guest without DAWR1
> > >      support (No warnings). Error out only if user tries with
> > >      cap-dawr1=on.
> > >    - KVM supports DAWR1: Boot the guest with DAWR1 support, unless
> > >      user specifies cap-dawr1=off.
> > >    - TCG guest: Ignore cap-dawr1 i.e. boot as if there is only
> > >      DAWR0 (Should be fixed in future while adding PowerPC watch-
> > >      point support in TCG mode)
> > > 
> > > Power10 predecessor behavior:
> > >    - KVM guest: Boot the guest without DAWR1 support. Error out
> > >      if user tries with cap-dawr1=on.
> > >    - TCG guest: Ignore cap-dawr1 i.e. boot as if there is only
> > >      DAWR0 (Should be fixed in future while adding PowerPC watch-
> > >      point support in TCG mode)
> > > 
> > > > I'm wondering if we're actually just better off setting the pa feature
> > > > just based on the guest CPU model.  TCG will be broken if you try to
> > > > use it, but then, it already is.  AFAIK there's no inherent reason we
> > > > couldn't implement DAWR support in TCG, it's just never been worth the
> > > > trouble.
> > > 
> > > Correct. Probably there is no practical usecase for DAWR in TCG mode.
> > 
> > What's the expected behavior ? Is it to generate a DSI if we have a DAWR
> > match ?
> 
> Yes. DSI is the main thing. But many auxiliary stuff, off the top of my
> head:
>  - DAR needs to be set. Now, DAR value is set differently on p8 vs p10
>    (not sure about p9 because there was hw bug and thus we needed to
>    fully disable DAWR on p9).
>  - DAWR matching criteria for quadword instruction are different for
>    p8/p9 vs p10.
>  - P10 supports 512 byte unaligned watchpoints but p8/p9 does not.
> 
> Kernel is aware of these differences and thus handles these scenarios,
> sometimes as special case. i.e. Qemu will need to mimic the exact hw
> behavior for the specific revision of processor.

I don't actually know if qemu has TCG watchpoint support on any
hardware.  Presumably it would mean instrumenting all the tcg loads
and stores.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]