On 25/11/2021 15:10, Hannes Reinecke wrote:
+/**
+ * scsi_get_internal_cmd - allocate an internal SCSI command
+ * @sdev: SCSI device from which to allocate the command
+ * @data_direction: Data direction for the allocated command
+ * @nowait: do not wait for command allocation to succeed.
+ *
+ * Allocates a SCSI command for internal LLDD use.
+ */
+struct scsi_cmnd *scsi_get_internal_cmd(struct scsi_device *sdev,
+ int data_direction, bool nowait)
+{
+ struct request *rq;
+ struct scsi_cmnd *scmd;
+ blk_mq_req_flags_t flags = 0;
+ int op;
+
+ if (nowait)
+ flags |= BLK_MQ_REQ_NOWAIT;
+ op = (data_direction == DMA_TO_DEVICE) ?
+ REQ_OP_DRV_OUT : REQ_OP_DRV_IN;
+ rq = blk_mq_alloc_request(sdev->request_queue, op, flags);
+ if (IS_ERR(rq))
+ return NULL;
+ scmd = blk_mq_rq_to_pdu(rq);
+ scmd->device = sdev;
+ return scmd;
+}
+EXPORT_SYMBOL_GPL(scsi_get_internal_cmd);
So there are a couple of generally-accepted grievances about this approach:
a. we're being allocated a scsi_cmnd, but not using what is being
allocated as a scsi_cmnd, but rather just a holder as a reference to an
allocated tag
b. we're being allocated a request, which is not being sent through the
block layer*
It just seems to me that what the block layer is providing is not suitable.
How about these:
a. allow block driver to specify size of reserved request PDU separately
to regular requests, so we can use something like this for rsvd commands:
struct scsi_rsvd_cmnd {
struct scsi_device *sdev;
}
And fix up SCSI iter functions and LLDs to deal with it.
b. provide block layer API to provide just same as is returned from
blk_mq_unique_tag(), but no request is provided. This just gives what we
need but would be disruptive in scsi layer and LLDs.
c. as alternative to b., send all rsvd requests through the block layer,
but can be very difficult+disruptive for users
*For polling rsvd commands on a poll queue (which we will need for
hisi_sas driver and maybe others for poll mode support), we would need
to send the request through the block layer, but block layer polling
requires a request with a bio, which is a problem.
Thanks,
John
