[PATCH v8 10/10] nvme: Atomic write support

[Kanchan Joshi]

On 6/10/2024 4:13 PM, John Garry wrote:
> +static bool nvme_valid_atomic_write(struct request *req)
> +{
> +	struct request_queue *q = req->q;
> +	u32 boundary_bytes = queue_atomic_write_boundary_bytes(q);
> +
> +	if (blk_rq_bytes(req) > queue_atomic_write_unit_max_bytes(q))
> +		return false;
> +
> +	if (boundary_bytes) {
> +		u64 mask = boundary_bytes - 1, imask = ~mask;
> +		u64 start = blk_rq_pos(req) << SECTOR_SHIFT;
> +		u64 end = start + blk_rq_bytes(req) - 1;
> +
> +		/* If greater then must be crossing a boundary */
> +		if (blk_rq_bytes(req) > boundary_bytes)
> +			return false;

Nit: I'd cache blk_rq_bytes(req), since that is repeating and this 
function is called for each atomic IO.

> +
> +		if ((start & imask) != (end & imask))
> +			return false;
> +	}
> +
> +	return true;
> +}
> +
>   static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
>   		struct request *req, struct nvme_command *cmnd,
>   		enum nvme_opcode op)
> @@ -941,6 +965,12 @@ static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
>   
>   	if (req->cmd_flags & REQ_RAHEAD)
>   		dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
> +	/*
> +	 * Ensure that nothing has been sent which cannot be executed
> +	 * atomically.
> +	 */
> +	if (req->cmd_flags & REQ_ATOMIC && !nvme_valid_atomic_write(req))
> +		return BLK_STS_INVAL;
>   

Is this validity check specific to NVMe or should this be moved up to 
block layer as it also knows the limits?