On 1/27/26 23:41, [email protected] wrote:
Alejandro Lucero Palau wrote:
[..]
I will take a look at this presentation, but I think there could be
another option where accelerators information is obtained during pci
enumeration by the kernel and using this information by this
functionality skipping those ranges allocated to them. Forcing them to
be compiled with the kernel would go against what distributions
currently and widely do with initramfs. Not sure if some current "early"
stubs could be used for this though but the information needs to be
recollected before this code does the checks.
The simple path is "do not use EFI_MEMORY_SP for accelerator memory".
Sure. That is what I hope our device will end up having ... since
neither hmem nor dax is an option for us.
However, if the accelerator wants to publish memory as EFI_MEMORY_SP
then it needs to coordinate with the kernel's default behavior somehow.
I think some Type2 drivers could be happy with dax and therefore using
EFI_MEMORY_SP, so yes, that is what I meant: there is another option
instead of forcing drivers to be present at the time of this decision.
If someone reading is working on Type2 drivers and see this
suitable/required, please tell. I'll be interested in doing it or helping.
That means expanding the list of drivers that dax_hmem needs to await
before it can make a determination, or teaching dax_hmem to look for a
secondary indication that it should never fall back to the default
behavior.
I think waiting could be problematic as some Type2 drivers could not be
automatically load. It looks like if a CXL region is not backing the
Type2 CXL.mem completely should not impact dax devices and cxl regions
maybe being used at Type2 driver probe. Would a warning be enough?
Talk to your AMD peers Paul and Rajneesh about their needs. I took it on
faith that the use case was required.
After reading that presentation, I think this is a different subject.
Assuming case 1 there is what you have in mind, and if I understand it
properly, that could be useful for companies owning the full platform,
but not sure adding a specific acpi driver per device makes sense for
less-powerful vendors. Anyway, I will talk with them as the memory
allocation part which seems to be one thing to do by those acpi drivers
is interesting.
Thank you
