Re: libnvme API design

[Christoph Hellwig <hch@lst.de>]

On Thu, Oct 09, 2025 at 01:07:47PM +0200, Daniel Wagner wrote:
> Hi Christoph,
> 
> As you have suggested the command API should be inverted, that is the
> only stable API will be nvme_submit_admin_passthru and struct
> nvme_passthru_cmd and a bunch of helpers to initialize the cmd.
> 
> I've done updated a bunch of the identify commands and below is the
> result. I think we could reduce the number of helpers a bit,
> 
>   nvme_nvm_init_identify_ns vs nvme_nvm_init_identify_ns_csi.
> 
> Is this what you had in mind

Yes.

> (obviously I haven't gone the whole way to
> macros only, I like some type safty)?

What actual type safety do we actually gain here?

> static inline
> void nvme_nvm_init_identify_ns(struct nvme_passthru_cmd *cmd,
> 			       __u32 nsid, struct nvme_id_ns *ns)
> {
> 	nvme_nvm_init_identify(cmd,
> 			       NVME_IDENTIFY_CNS_NS,
> 			       ns, sizeof(*ns));
> 	cmd->nsid = nsid;

Given that identify includes nsid and just specified how it is
cleared to zero for the non-ns commands, maybe pass a nsid to
nvme_nvm_init_identify to simplify things?

> 	cmd->cdw11 |=  NVME_SET(nvmsetid, IDENTIFY_CDW11_CNSSPECID);

And NVME_SET still feels very oddly name..