On 02/06/2014 04:11 PM, Nicholas A. Bellinger wrote:
>> +struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
>> > +{
>> > +	struct Scsi_Host *shost = sdev->host;
>> > +	struct blk_mq_hw_ctx *hctx;
>> > +	struct request_queue *q;
>> > +	struct request *rq;
>> > +	struct scsi_cmnd *cmd;
>> > +	struct blk_mq_reg reg;
>> > +	int i, j, sgl_size;
>> > +
>> > +	memset(&reg, 0, sizeof(reg));
>> > +	reg.ops = &scsi_mq_ops;
>> > +	reg.queue_depth = shost->cmd_per_lun;
>> > +	if (!reg.queue_depth)
>> > +		reg.queue_depth = 1;
>> > +
>> > +	/* XXX: what to do about chained S/G lists? */
>> > +	if (shost->hostt->sg_tablesize > SCSI_MAX_SG_SEGMENTS)
>> > +		shost->sg_tablesize = SCSI_MAX_SG_SEGMENTS;
>> > +	sgl_size = shost->sg_tablesize * sizeof(struct scatterlist);
>> > +
>> > +	reg.cmd_size = sizeof(struct scsi_cmnd) +
>> > +			sgl_size +
>> > +			shost->hostt->cmd_size;
>> > +	if (scsi_host_get_prot(shost))
>> > +		reg.cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
> OK, so your in-lining the allocation of data + protection SGLs from
> blk-mq..
> 
> The original prototype code was doing these allocations separately below
> for each pre-allocated cmd, and offering LLD's to optionally
> pre-allocate their own descripts using sh->hostt->cmd_size if
> necessary..
> 
> This was necessary to eliminate all fast-path allocations for
> virtio-scsi, and I'd like to see something similar here as an optional
> feature as well.

Yeah, it would be nice if like in Nick's patches, the driver could just
set the scsi_host_template->cmd_size then when the scsi_cmnd got to the
driver's queuecommand, the driver could just get its internal cmd struct
from the scsi_cmnd struct (for example in Nick's patch it was off the
SCp.ptr).

I started converting my iscsi mq patch from Nick's code to Christoph's
and am currently trying to figure out how to setup the
scsi_host_template->cmd_pool.

Current iscsi patch is attached if anyone cares.

However, one question I had with both approaches is how to deal with per
cmd pci/dma memory and preallocations.