Re: [PATCH v5 05/19] arm64: Add support for trace synchronization barrier

[Catalin Marinas]

On Wed, Mar 24, 2021 at 05:06:58PM +0000, Suzuki K Poulose wrote:
> On 24/03/2021 16:30, Marc Zyngier wrote:
> > On Wed, 24 Mar 2021 16:25:12 +0000,
> > Suzuki K Poulose <suzuki.poulose@xxxxxxx> wrote:
> > > 
> > > On 24/03/2021 16:16, Marc Zyngier wrote:
> > > > On Wed, 24 Mar 2021 15:51:14 +0000,
> > > > Suzuki K Poulose <suzuki.poulose@xxxxxxx> wrote:
> > > > > 
> > > > > On 24/03/2021 13:49, Marc Zyngier wrote:
> > > > > > On Wed, 24 Mar 2021 09:39:13 +0000,
> > > > > > Suzuki K Poulose <suzuki.poulose@xxxxxxx> wrote:
> > > > > > > 
> > > > > > > On 23/03/2021 18:21, Catalin Marinas wrote:
> > > > > > > > Hi Suzuki?
> > > > > > > > 
> > > > > > > > On Tue, Mar 23, 2021 at 12:06:33PM +0000, Suzuki K Poulose wrote:
> > > > > > > > > tsb csync synchronizes the trace operation of instructions.
> > > > > > > > > The instruction is a nop when FEAT_TRF is not implemented.
> > > > > > > > > 
> > > > > > > > > Cc: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>
> > > > > > > > > Cc: Mike Leach <mike.leach@xxxxxxxxxx>
> > > > > > > > > Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
> > > > > > > > > Cc: Will Deacon <will.deacon@xxxxxxx>
> > > > > > > > > Signed-off-by: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> > > > > > > > 
> > > > > > > > How do you plan to merge these patches? If they go via the coresight
> > > > > > > > tree:
> > > > > > > > 
> > > > > > > 
> > > > > > > Ideally all of this should go via the CoreSight tree to have the
> > > > > > > dependencies solved at one place. But there are some issues :
> > > > > > > 
> > > > > > > If this makes to 5.13 queue for CoreSight,
> > > > > > > 
> > > > > > > 1) CoreSight next is based on rc2 at the moment and we have fixes gone
> > > > > > > into rc3 and later, which this series will depend on. (We could move
> > > > > > > the next tree forward to a later rc to solve this).
> > > > > > > 
> > > > > > > 2) There could be conflicts with the kvmarm tree for the KVM host
> > > > > > > changes (That has dependency on the TRBE definitions patch).
> > > > > > > 
> > > > > > > If it doesn't make to 5.13 queue, it would be good to have this patch,
> > > > > > > the TRBE defintions and the KVM host patches queued for 5.13 (not sure
> > > > > > > if this is acceptable) and we could rebase the CoreSight changes on 5.13
> > > > > > > and push it to next release.
> > > > > > > 
> > > > > > > I am open for other suggestions.
> > > > > > > 
> > > > > > > Marc, Mathieu,
> > > > > > > 
> > > > > > > Thoughts ?
> > > > > > 
> > > > > > I was planning to take the first two patches in 5.12 as fixes (they
> > > > > > are queued already, and would hopefully land in -rc5). If that doesn't
> > > > > > fit with the plan, please let me know ASAP.
> > > > > 
> > > > > Marc,
> > > > > 
> > > > > I think it would be better to hold on pushing those patches until we
> > > > > have a clarity on how things will go.
> > > > 
> > > > OK. I thought there was a need for these patches to prevent guest
> > > > access to the v8.4 self hosted tracing feature that went in 5.12
> > > > though[1]... Did I get it wrong?
> > > 
> > > Yes, that is correct. The guest could access the Trace Filter Control
> > > register and fiddle with the host settings, without this patch.
> > > e.g, it could disable tracing at EL0/EL1, without the host being
> > > aware on nVHE host.
> > 
> > OK, so we definitely do need these patches, don't we? Both? Just one?
> > Please have a look at kvmarm/fixes and tell me what I must keep.
> 
> Both of them are fixes.
> 
> commit "KVM: arm64: Disable guest access to trace filter controls"
>  - This fixes guest fiddling with the trace filter control as described
> above.
> 
> commit "KVM: arm64: Hide system instruction access to Trace registers"
>  - Fixes the Hypervisor to advertise what it doesn't support. i.e
>    stop advertising trace system instruction access to a guest.
>    Otherwise a guest which trusts the ID registers
>    (ID_AA64DFR0_EL1.TRACEVER == 1) can crash while trying to access the
>    trace register as we trap the accesses (CPTR_EL2.TTA == 1). On Linux,
>    the ETM drivers need a DT explicitly advertising the support. So,
>    this is not immediately impacted. And this fix goes a long way back
>    in the history, when the CPTR_EL2.TTA was added.
> 
> Now, the reason for asking you to hold on is the way this could create
> conflicts in merging the rest of the series.

The way we normally work around this is to either rebase your series on
top of -rc5 when the fixes go in or, if you want an earlier -rc base,
Marc can put them on a stable branch somewhere that you can use.

In the worst case you can merge the patches twice but that's rarely
needed.

-- 
Catalin