Update COORDINATION.md with Marvell review comments

[sburla-marvell]

Add Marvell comments and CRS details

[glimchb]

@ballle98 @RezaBacchus @dandaly @Gal-Zaidman please review

[dandaly]

Hi, I can merge the option 2 feedback into another PR:

• I will add that Driver Ready Check works for Multi-Host
• I will add as a Pro that this solution already works without interdependencies with other standards bodies (PCI, DMTF, etc.).
• I will add that it would make it easier to adopt if we could standardize the virtio solutions (either a standard OROM signaling to the device, or adopt the approaches from NVMe and idpf). For option 2 did I catch it all?

[RezaBacchus]

Greetings Dan: The abstraction you proposed is highly desirable, however, OROMs run at the end of POST and this conflicts with the need for the OROM to run when the card is discovered, so that it may stall UEFI until all the Bus/Dev/Fun on the card are discoverable. Re writing UEFI to make the OROM run when the card is discovered is a heavy lift and risky because the system is not stable during descovery.

[sburla-marvell]

Thanks @sburla-marvell great comments. Do yiu want to merge this content or just collect feedback? Let me know

Most of it is just feedback. I can create a separate pull request for the CRS part if needed.

[glimchb]

Most of it is just feedback. I can create a separate pull request for the CRS part if needed.

yes, please do, so I can merge the CRS part

[dandaly]

Greetings Dan: The abstraction you proposed is highly desirable, however, OROMs run at the end of POST and this conflicts with the need for the OROM to run when the card is discovered, so that it may stall UEFI until all the Bus/Dev/Fun on the card are discoverable. Re writing UEFI to make the OROM run when the card is discovered is a heavy lift and risky because the system is not stable during descovery.

Hi Reza, This OROM solution solves the coordination problem between a PCI compatible device and the host. If the device is not PCI compatible, you need some other solution. This driver based solution is working today, it doesn't require any UEFI changes. I think we need to split Coordination between solving these two separate problems:

• Coordinating between host and infrastructure device within the confines of existing specs and standard solutions (PCI, UEFI, BIOS, all unchanged). Make it work without any assumptions.

• Propose changes to specs and standard solutions to workaround specific issues, for example if certain DPUs cannot be PCI compliant

  This way we don't gate IPU/DPU provisioning on requiring a new PCI, UEFI or BIOS change in order for it to work the OPI way. We can also specify the OPI way to workaround specific incompatibilities, whatever they may be.

[glimchb]

This way we don't gate IPU/DPU provisioning on requiring a new PCI, UEFI or BIOS change in order for it to work the OPI way. We can also specify the OPI way to workaround specific incompatibilities, whatever they may be.

we should try to find a way that solves all vendors. if we can't, we can't... but we need to try first...

[dandaly]

This way we don't gate IPU/DPU provisioning on requiring a new PCI, UEFI or BIOS change in order for it to work the OPI way. We can also specify the OPI way to workaround specific incompatibilities, whatever they may be.

we should try to find a way that solves all vendors. if we can't, we can't... but we need to try first...

I propose we answer the OPI provisioning question within the constraints of what's existing, and see where we land. I also haven't heard from any other vendor that they can't work within the PCI spec, or that they can't work in existing servers.

[glimchb]

I also haven't heard from any other vendor that they can't work within the PCI spec, or that they can't work in existing servers.

I got feedback from one DPU vendor that strongly advocates for Option3 with OOB management via BMC in order not to rely on PCIe timing. Even if they are PCIe compliant card.

[dandaly]

I got feedback from one DPU vendor that strongly advocates for Option3 with OOB management via BMC in order not to rely on PCIe timing. Even if they are PCIe compliant card.

Can we get this rationale into the document? We need to have a clear reason why we need to create a dependency like what Option 3 proposes, since that dependency will limit adoption.