scsi: core: Remove an atomic instruction from the hot path

From scsi_init_command(), a function called by scsi_mq_prep_fn():

/* zero out the cmd, except for the embedded scsi_request */
memset((char *)cmd + sizeof(cmd->req), 0,
sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);

In other words, scsi_mq_prep_fn() clears scsi_cmnd.flags. Hence move the
clear_bit() call into the else branch, the only branch in which this code
is necessary.

See also commit f1342709d18a ("scsi: Do not rely on blk-mq for double
completions").

Cc: Keith Busch <[e-mail]>
Cc: Christoph Hellwig <[e-mail]>
Cc: Hannes Reinecke <[e-mail]>
Signed-off-by: Bart Van Assche <[e-mail]>
Reviewed-by: Keith Busch <[e-mail]>
Signed-off-by: Martin K. Petersen <[e-mail]>

drivers/scsi/scsi_lib.c

···
 	if (!scsi_host_queue_ready(q, shost, sdev))
 		goto out_dec_target_busy;
 
-	clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
 	if (!(req->rq_flags & RQF_DONTPREP)) {
 		ret = scsi_mq_prep_fn(req);
 		if (ret != BLK_STS_OK)
 			goto out_dec_host_busy;
 		req->rq_flags |= RQF_DONTPREP;
 	} else {
+		clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
 		blk_mq_start_request(req);
 	}