Re: [PATCH v2 0/5] Address translation for HDM decoding

[Alison Schofield]

On Wed, Jul 10, 2024 at 05:02:46PM -0700, Alison Schofield wrote:
> On Mon, Jul 01, 2024 at 07:47:48PM +0200, Robert Richter wrote:
> > Default expectation of Linux is that HPA == SPA, which means that
> > hardware addresses in the decoders are the same as the kernel sees
> > them. However, there are platforms where this is not the case and an
> > address translation between decoder's (HPA) and the system's physical
> > addresses (SPA) is needed.
> > 
> > This series implements address translation for HDM decoding. The
> > implementation follows the rule that the representation of hardware
> > address ranges in the kernel are all SPA. If decoder registers (HDM
> > decoder cap or register range) are not SPA, a base offset must be
> > applied. Translation happens when accessing the registers back and
> > forth. After a read access an address will be converted to SPA and
> > before a write access the programmed address is translated from an
> > SPA. The decoder register access can be easily encapsulated by address
> > translation and thus there are only a few places where translation is
> > needed and the code must be changed. This is implemented in patch #2,
> > patch #1 is a prerequisite.
> > 
> > Address translation is restricted to platforms that need it. As such a
> > platform check is needed and a flag is introduced for this (patch #3).
> > 
> > For address translation the base offset must be determined for the
> > memory domain. Depending on the platform there are various options for
> > this. The address range in the CEDT's CFWMS entry of the CXL host
> > bridge can be used to determine the decoder's base address (patch
> > #4). This is enabled for AMD Zen4 platforms (patch #5).
> 
> 
> Hi Robert,
> 
> This HPA->SPA work needs to be done for addresses reported directly by
> devices, ie DPAs in poison and other events - right?
> 
> For the XOR case, we discover the need for HPA->SPA while parsing the
> CFMWS and add a cxl_hpa_to_spa_fn to the struct cxl_root_decoder. Later,
> when the driver wants to translate a DPA (to include in a TRACE_EVENT)
> it uses that 'extra mile' HPA->SPA function.
> 
> See Patch 2 in this series:
> https://lore.kernel.org/cover.1719980933.git.alison.schofield@xxxxxxxxx/T/#m9206e1f872ef252dbb54ce7f0365f0b267179fda

Better link:
https://lore.kernel.org/linux-cxl/cover.1719980933.git.alison.schofield@xxxxxxxxx/

> 
> It seems the Zen4 extra mile is a simple offset from the base calc.
> Do you think a zen4 hpa->spa function will fit in with what I've done?
> 
> FWIW I took this code for a spin through cxl-test on it's own and combined
> w the xor address tranlation patch set and no collisions, all humming along
> nicely (for non-zen config).
> 
> --Alison
> 
> > 
> snip
> > 
>