[PATCH] nvmet-loop: use nr_phys_segments when map rq to sgl

[hch@infradead.org (Christoph Hellwig)]

On Mon, May 14, 2018@03:17:51PM -0700, chaitany kulkarni wrote:
> +static inline void nvme_setup_write_zeroes(struct nvme_ns *ns,
> +               struct request *req, struct nvme_command *cmnd)
> +{
> +       struct nvme_write_zeroes_cmd *write_zeroes = &cmnd->write_zeroes;
> +
> +       memset(cmnd, 0, sizeof(*cmnd));
> +       write_zeroes->opcode = nvme_cmd_write_zeroes;
> +       write_zeroes->nsid = cpu_to_le32(ns->head->ns_id);
> +       write_zeroes->slba =
> +               cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
> +       write_zeroes->length =
> +               cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
> +       write_zeroes->control = 0;
> +}

This doesn't even set RQF_SPECIAL_PAYLOAD, so that is not the issue.

Actually reviewing things again I think your original patch is fine
as it actually uses blk_rq_nr_phys_segments() which checks
RQF_SPECIAL_PAYLOAD.