Re: [PATCH v3 00/10] arm64: ptdump: View the second stage page-tables

From: Oliver Upton
Date: Wed Nov 22 2023 - 18:18:56 EST


Hi Seb,

On Wed, Nov 15, 2023 at 05:16:30PM +0000, Sebastian Ene wrote:
> Hi,
>
> This can be used as a debugging tool for dumping the second stage
> page-tables.
>
> When CONFIG_PTDUMP_STAGE2_DEBUGFS is enabled, ptdump registers
> '/sys/debug/kvm/<guest_id>/stage2_page_tables' entry with debugfs
> upon guest creation. This allows userspace tools (eg. cat) to dump the
> stage-2 pagetables by reading the registered file.
>
> Reading the debugfs file shows stage-2 memory ranges in following format:
> <IPA range> <size> <descriptor type> <access permissions> <mem_attributes>
>
> Under pKVM configuration(kvm-arm.mode=protected) ptdump registers an entry
> for the host stage-2 pagetables in the following path:
> /sys/debug/kvm/host_stage2_page_tables/
>
> The tool interprets the pKVM ownership annotation stored in the invalid
> entries and dumps to the console the ownership information. To be able
> to access the host stage-2 page-tables from the kernel, a new hypervisor
> call was introduced which allows us to snapshot the page-tables in a host
> provided buffer. The hypervisor call is hidden behind CONFIG_NVHE_EL2_DEBUG
> as this should be used under debugging environment.

While I think the value of the feature you're proposing is great, I'm
not a fan of the current shape of this series.

Reusing note_page() for the stage-2 dump is somewhat convenient, but the
series pulls a **massive** amount of KVM details outside of KVM:

- Open-coding the whole snapshotting interface with EL2 outside of KVM.
This is a complete non-starter for me; the kernel<->EL2 interface
needs to be owned by the EL1 portions of KVM.

- Building page-table walkers using the KVM pgtable library outside of
KVM.

- Copying (rather than directly calling) the logic responsible for
things like FWB and PGD concatenation.

- Hoisting the definition of _software bits_ outside of KVM. I'm less
concerned about hardware bits since they have an unambiguous meaning.

I think exporting the necessary stuff from ptdump into KVM will lead to
a much cleaner implementation.

--
Thanks,
Oliver