On Fri, Jul 9, 2021 at 8:24 AM Jeff Moyer <[email protected]> wrote: > > Dan Williams <[email protected]> writes: > > > [ add Jeff, Michal, and Adam ] > > [ adding Bryan Gurney, who is helping out with RHEL packaging ] > > > Hey ndctl distro maintainers, > > > > Just wanted to highlight this new tool submission for your > > consideration. The goal here is to have a Linux native provisioning > > tool that covers the basics of the functionality that is outside of > > the ACPI specification, and reduce the need for ipmctl outside of > > exceptional device-specific debug scenarios. Recall that the ACPI NFIT > > communicates the static region configuration to the OS, but changing > > that configuration is a device-specific protocol plus a reboot. Until > > the arrival of pcdctl, region provisioning required ipmctl. > > It's great to see progress on this, thanks! Shipping another utility as > part of the ndctl package is fine with me, though I'm not sure why we > wouldn't just make this an ndctl sub-command. From a user's > perspective, these are all operations on or about nvdimms. ipmctl > didn't have separate utilities for provisioning goals and namespace > configuration, for example.
True, but ipmctl also did not make an attempt to support anything other than Intel devices, and later versions abandoned the namespace setup code in favor of "native OS" capabilities (ndctl on Linux). The main rationale for splitting region provisioning to dedicated tooling is the observation that region provisioning semantics are platform specific. It is already the case that IBM devices have their own provisioning tool with different semantics for the "PAPR" family. CXL region provisioning semantics again are much different than what is done for DDR-T devices (see below). So rather than try to abstract all that under ndctl that wants to be vendor agnostic, offload that to platform specific tools. My hope is that more tools like this do not proliferate as the industry unifies on common standards for persistent memory like CXL. That said, the new commands could be placed under a vendor/platform-specific name in ndctl, like: ndctl list-ipm-region ndctl reconfigure-ipm-region ...just not my first choice given the success to date of keeping vendor details out of the command line interface of ndctl. The primary blocker for ndctl to generic region provisioning would be a kernel driver model for it, but I don't know how to reconcile "ipm-regions" requiring a reboot and a BIOS validation step vs buses like CXL that can reconfigure interleave sets at runtime. > > I will note that CXL moves the region configuration into the base CXL > > specification so the ndctl project will pick up a "cxl-cli" tool for > > that purpose. In general, the ndctl project is open to carrying > > support for persistent memory devices with open specifications. In > > this case the provisioning specification for devices formerly driven > > by ipmctl was opened up and provided here: > > Is there a meaningful difference to the user? Can you show some > examples of how configuration would be different between cxl-attached > pmem and memory-bus attached pmem? Yes, CXL exposes several more details and degrees of freedom to system software. Before I list those I'll point out that to keep pcdctl simple it only handles the simple / common configurations: all performance-pmem (interleaved), all fault tolerant pmem (non-interleaved), all volatile with memory-side-caching. Any custom/expert configuration outside of those common cases is punted to ipmctl. In comparison, the CXL tool will need to handle the full range of configuration complexity. The main difference to end users when provisioning regions on CXL is the wider range of resources they need to consider. The CXL specific resources include: - Available PMEM capable address space as described by the ACPI CFMWS - Device performance that matches the address space traffic class - Decoder resources at each level of the hierarchy. I.e. a device may be able to participate in 4 different interleave configurations, but depending on the switch topology upstream of that device it may be constrained to a smaller set. - Volatile memory vs PMEM partitioning on the device. The NVDIMM sub-system and ndctl will not have any responsibility for the volatile memory side of CXL. To me that looks like sufficient complexity to warrant a dedicated CXL tool rather than try to find a lowest-common-denominator abstraction that melds with ipm-regions for ndctl to drive generically. The CXL tool will also handle firmware update and other CXL generic functionality outside of PMEM. > > https://cdrdv2.intel.com/v1/dl/getContent/634430 > > > > Please comment on its suitability for shipping in distros alongside > > the ndctl tool. > > It's completely fine to ship more tools with ndctl. I would like a > better overall picture of configuration from the admin's perspective. > At first glance, I think we're adding unneeded complexity. You mean the complexity of having to determine which platform region provisioning tool to use before you can use ndctl to do the rest? > > Cheers, > Jeff > > p.s. I don't find the name 'pdctl' particularly endearing. If we do > stick with a separate utility, I'd suggest coming up with a more > descriptive name. How about "ipmregion"? Where "ipm" is already in the wild as an identifier for DDR-T configuration, and unlike ipmctl it only handles the region provisioning subset?
