On 2021-08-16 11:21, Ard Biesheuvel wrote:

On Mon, 16 Aug 2021 at 11:59, Lorenzo Pieralisi
[off-list ref] wrote:

quoted

On Wed, Aug 11, 2021 at 03:55:08PM +0100, Lorenzo Pieralisi wrote:

quoted

On Wed, Aug 11, 2021 at 03:08:24PM +0100, Christoph Hellwig wrote:

quoted

On Wed, Aug 11, 2021 at 12:40:28PM +0200, Ard Biesheuvel wrote:

quoted

The whole problem we are solving here is that ACPI, being based on
x86, conflates MMIO mappings with memory mappings, and has been using
the same underlying infrastructure for either.

So let's fix that problem instead of papering over it.

Patch (3) in this series is a fix - I would ask whether it makes
sense to merge patches (2-3) now and think about reworking the current
ACPI IO/MEM mapping API later, it can be an invasive change for a fix,
assuming we agree on how to rework the ACPI IO/MEM mapping API.

What should we do then with this series ?

It is not even clear that reworking the ACPI core is feasible to begin
with, OTOH, fixing a sparse warning is arguably not a critical bug fix
either, so I'd suggest we just drop that bit.

Indeed, the only way to truly fix the issue is to fire up the time 
machine and rewrite the ACPI and EFI specs to not define that tables and 
data may or may not be required to be mapped as Device memory depending 
on the whims of the firmware. Otherwise we're basically always going to 
have one or more casts *somewhere*, even if we were to play it safe and 
return everything as iomem instead.

I guess for read-only access to tables, the core code might be able to 
maintain a shadow copy of anything device-memory-mapped in normal memory 
and expose that instead, but if anything has to be writeable I'm not 
sure how we could abstract that "properly".

Robin.