Re: [PATCH v2 1/5] MdeModulePkg/CapsulePei: clean Dcache before consuming capsule data

[Ard Biesheuvel <ard.biesheuvel@linaro.org>]

On 11 June 2018 at 23:40, Kinney, Michael D <michael.d.kinney@intel.com> wrote:
> Hi Ard,
>
> I would still prefer the cache maintenance be done independent
> of capsules in case there is other memory ranges that need to
> be flushed for other features.
>
> The CacheMaintenanceLib for ARM has several services that are
> not implemented and ASSERT().  I found some ARM documentation
> that says that caches can be flushed by looping through the
> indexes.
>
> http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0201d/ch03s03s05.html
>

Please don't quote severely outdated documentation out of context.
ARM9 is a ~15 year old core that predates SMP on ARM by several years.

> implementations that may not work with that algorithm.  However
> that could require a customized CacheMaintenanceLib.  Can you
> provide the flush all using the documented algorithm and use that
> instance for systems that are compatible with the documented algorithm.

None of the ARM cores we support in EDK2 today (ARMv7 and later) are
compatible with the documented algorithm.

> Then, the flush all API can be used in ResetSystem() or SEC Phase
> when the boot mode is warm reset.
>

Flushing the entire cache is simply not possible. The set/way
operations are only intended to be used before the core enters the
coherency domain or after it leaves it again (i.e., when cores are
powered up or down). Those set/way operations are not broadcast to
other cores (or other masters such as DMA capable peripherals with
caches), which means that a cache line that is cleaned by set/way may
simply end up in another cache and never make it to main memory. In
other words, set/way operations manage the cache itself rather than
the contents of the caches.

The only way to perform cache maintenance in a way to ensure that the
*content* makes it to main memory is to use cache maintenance by
virtual address. This requires that you know the virtual address to
begin with, and obviously requires that a mapping exists for that
virtual address when executed with the MMU on.

The bottom line is that there is no flush all API, and we will have to
work around that (and believe me, this is not the first time we are
struggling to deal with this limitation).

So to summarize again, we have the following options,
- UpdateCapsule - problematic because it may be called at runtime and
the firmware has no means of translating the physical addresses
- ResetSystem - same as above: it is a runtime service, and so it
cannot rely on a mapping to exist for those physical addresses
- SEC - lacks the information about where the capsule resides
- CapsulePei - already extracts the information about the capsule
address in memory, and can perform the cache maintenance right before
consuming the data.