On Fri, 07 Sep 2007 17:36:47 -0500
Mike Christie [off-list ref] wrote:

quoted

+/*
+ * map SG list
+ */
+static int bnx2i_map_sg(struct bnx2i_hba *hba, struct bnx2i_cmd *cmd)
+{
+	struct scsi_cmnd *sc = cmd->scsi_cmd;
+	struct iscsi_bd *bd = cmd->bd_tbl->bd_tbl;
+	struct scatterlist *sg;
+	int byte_count = 0;
+	int sg_frags;
+	int bd_count = 0;
+	int sg_count;
+	int sg_len;
+	u64 addr;
+	int i;
+
+	sg = sc->request_buffer;
+	sg_count = pci_map_sg(hba->pci_dev, sg, sc->use_sg,
+			      sc->sc_data_direction);

Can you use scsi_dma_map() here?

quoted

+	for (i = 0; i < sg_count; i++) {
+		sg_len = sg_dma_len(sg);
+		addr = sg_dma_address(sg);
+		if (sg_len > MAX_BD_LENGTH)
+			sg_frags = bnx2i_split_bd(cmd, addr, sg_len,
+						  bd_count);

Please use scsi_for_each_sg().

You can't directly access to use_sg, request_buffer, request_bufflen,
and resid in scsi_cmnd structure. Please use the scsi data accessors
in scsi_cmnd.h: scsi_sg_count, scsi_sglist, scsi_bufflen,
scsi_set_resid, and scsi_get_resid.

If you call blk_queue_max_segment_size() in the slave_configure callout 
you can limit the size of the segments that the block layer builds so 
they are smaller than MAX_BD_LENGTH. However, I am not sure how useful 
that is. I think DMA-API.txt states that the mapping code is ok to 
merged mutliple sglists entries into one so I think that means that we 
can still end up with an entry that is larger than MAX_BD_LENGTH. Not 
sure if there is way to tell the pci/dma map_sg code to limit this too.

Yeah, iommu code ignores the lld limitations (the problem is that the
lld limitations are in request_queue and iommu code can't access to
request_queue). There is no way to tell iommu code about the lld
limitations.