On 5/27/22 1:29 AM, Kanchan Joshi wrote:
> On Thu, May 26, 2022 at 08:18:04PM +0530, Ankit Kumar wrote:
>
>> +void fio_nvme_uring_cmd_prep(struct nvme_uring_cmd *cmd, struct io_u *io_u,
>> + struct iovec *iov)
>> +{
>> + struct nvme_data *data = FILE_ENG_DATA(io_u->file);
>> + __u64 slba;
>> + __u32 nlb;
>> +
>> + slba = io_u->offset / data->lba_size;
>> + nlb = (io_u->xfer_buflen / data->lba_size) - 1;
>> +
>> + memset(cmd, 0, sizeof(struct nvme_uring_cmd));
>
> Is this better or setting remaining fields (which are not populated
> down) to zero.
Since lba_size is a power of 2, it would be a lot more efficient to take
this division by non-constant out of the fast path and init a shift
value:
data->lba_shift = log2(data->lba_size);
and change these to:
slba = io_u->offset >> data->lba_shift;
and ditto for nlb.
>> + /* cdw10 and cdw11 represent starting lba */
>> + cmd->cdw10 = slba & 0xffffffff;
>> + cmd->cdw11 = slba >> 32;
>> + /* cdw12 represent number of lba's for read/write */
>> + cmd->cdw12 = nlb;
>> + if (iov) {
>> + iov->iov_base = io_u->xfer_buf;
>> + iov->iov_len = io_u->xfer_buflen;
>> + cmd->addr = (__u64)(uintptr_t)iov;
>> + cmd->data_len = 1;
>
> Is this correct? Do we always get 1 vector to deal with.
Yes
--
Jens Axboe
