Updates on storage standards

As he has in some previous editions of the Linux Storage, Filesystem, Memory-Management, and BPF Summit (LSFMM+BPF), Fred Knight gave an update on the status of various storage standards this year. In it, he looked at changes to the NVM Express (NVMe) standards in some detail. He also updated attendees on the fairly small changes that have come to the SCSI (T10) and ATA (T13) standards over the last few years.

He began with a bit of NVMe history. In May 2021, the 2.0 base specification was released, which split out several pieces of the standard into their own specifications. In addition to the base specification, there were three command-set specifications, three transport specifications, and the management-interface specification all released at the same time. Over the next three years, several revisions were released (2.0a-2.0e) that contained "corrections and clarifications" without any new features.

In August 2024, NVMe 2.1 was released, which incorporated those revisions and added new features. Several of the command-set specifications were revised in the process and two more were added: the computational programs command set and, going along with it, the subsystem local memory command set. The transport and management-interface specifications were revised, as well. A new boot specification was added to describe how NVMe would interface with EFI and the operating systems in order to boot over the network.

The 2.2 base specification had just been ratified earlier in March, Knight said. It had updates for several command-set specifications, the PCIe transport specification, and the boot specification.

Since the time period from August to March was not all that long, he said, there are not a lot of changes in 2.2; there are just a few technical proposals (TPs) and clarifications that were approved. TP 4194 for the base specification was aimed at better describing the scatter-gather list (SGL) feature; "the goal of this TP was not to change anything" but to make some of the assumptions about the interpretation of the existing specification more concrete. In particular, the validation of entries in an SGL only needs to be done for those entries that are actually used in the transfer, not all that are on the list. In addition, the errors that should be returned for various conditions are now fully described; "there could have been some confusion on how people interpreted which error is returned".

Engineering change notes (ECNs) tend to be clarifications to the specifications, while the TPs are functional changes, he said. The SGL changes added "should" and "shall" language, so those were considered functional changes. Several clarifications to the base specification came with the adoption of ECN 123. The clarification on the interaction between reservations and security send/receive is one that he thought might be of particular interest. The process for reading the boot partition has also changed to eliminate some timing-related problems with some devices.

In the zoned-namespace (ZNS) specification and several others, there is a change to specify all of the opcodes in a command set. Earlier, two commands ended up with the same opcode in the subsystem-local-memory (SLM) specification, which required breaking compatibility by changing the SLM opcode. In order to avoid that kind of problem in the future, ECN 125 changes all of the command sets to list all of the opcodes so there can be no mistaken duplication. In the PCIe transport specification, the eye-opening-measurement feature, which evaluates the signal quality on the bus, that had been added in 2.1 specification was changed for 2.2 to expand the eye data length (EDLEN) field from 16 to 32 bits. The boot specification was clarified to ensure that all of the universally unique IDs (UUIDs) used are specified in all lower case as part of ECN 123.

The T10 committee has been making a few changes to the SCSI specifications, Knight said. There has been work on the command duration limit (CDL) feature and on depopulation. The addition of DMTF security protocols and data models (SPDM) support is perhaps the most interesting part, he said. It will allow authentication between the host and the device, "so you can establish a secure channel and know that you're talking to the right device, not an impostor". The same set of features is being worked on for the ATA command-set (ACS) standards, he said, though there is also some work being done on resetting all of the write pointers at once for zoned-storage devices.

Knight closed by noting that the increased frequency of NVMe releases is intentional. More than three years elapsed between 2.0 and 2.1, so there were a lot of changes that required a great deal of work for the community to put it together. So, it has been decided that the committee will be releasing smaller, more frequent versions, "so that they're easier to digest, easier to understand, and easier to follow". The 2.2 release is the first and the hope is that more will be coming soon.