block: move dma drain handling to scsi

Don't burden the common block code with with specifics of the libata DMA
draining mechanism.  Instead move most of the code to the scsi midlayer.

That also means the nr_phys_segments adjustments in the blk-mq fast path
can go away entirely, given that SCSI never looks at nr_phys_segments
after mapping the request to a scatterlist.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

2 files changed, 44 insertions, 23 deletions

diff --git a/drivers/ata/libata-scsi.c b/drivers/ata/libata-scsi.c
index 36e588d88b95..feb13b8f93d7 100644
--- a/drivers/ata/libata-scsi.c
+++ b/drivers/ata/libata-scsi.c
@@ -1017,16 +1017,11 @@ void ata_scsi_sdev_config(struct scsi_device *sdev)
  *	RETURNS:
  *	1 if ; otherwise, 0.
  */
-static int atapi_drain_needed(struct request *rq)
+bool ata_scsi_dma_need_drain(struct request *rq)
 {
-	if (likely(!blk_rq_is_passthrough(rq)))
-		return 0;
-
-	if (!blk_rq_bytes(rq) || op_is_write(req_op(rq)))
-		return 0;
-
 	return atapi_cmd_type(scsi_req(rq)->cmd[0]) == ATAPI_MISC;
 }
+EXPORT_SYMBOL_GPL(ata_scsi_dma_need_drain);
 
 int ata_scsi_dev_config(struct scsi_device *sdev, struct ata_device *dev)
 {
@@ -1039,21 +1034,21 @@ int ata_scsi_dev_config(struct scsi_device *sdev, struct ata_device *dev)
 	blk_queue_max_hw_sectors(q, dev->max_sectors);
 
 	if (dev->class == ATA_DEV_ATAPI) {
-		void *buf;
-
 		sdev->sector_size = ATA_SECT_SIZE;
 
 		/* set DMA padding */
 		blk_queue_update_dma_pad(q, ATA_DMA_PAD_SZ - 1);
 
-		/* configure draining */
-		buf = kmalloc(ATAPI_MAX_DRAIN, q->bounce_gfp | GFP_KERNEL);
-		if (!buf) {
+		/* make room for appending the drain */
+		blk_queue_max_segments(q, queue_max_segments(q) - 1);
+
+		sdev->dma_drain_len = ATAPI_MAX_DRAIN;
+		sdev->dma_drain_buf = kmalloc(sdev->dma_drain_len,
+				q->bounce_gfp | GFP_KERNEL);
+		if (!sdev->dma_drain_buf) {
 			ata_dev_err(dev, "drain buffer allocation failed\n");
 			return -ENOMEM;
 		}
-
-		blk_queue_dma_drain(q, atapi_drain_needed, buf, ATAPI_MAX_DRAIN);
 	} else {
 		sdev->sector_size = ata_id_logical_sector_size(dev->id);
 		sdev->manage_start_stop = 1;
@@ -1135,7 +1130,6 @@ EXPORT_SYMBOL_GPL(ata_scsi_slave_config);
 void ata_scsi_slave_destroy(struct scsi_device *sdev)
 {
 	struct ata_port *ap = ata_shost_to_port(sdev->host);
-	struct request_queue *q = sdev->request_queue;
 	unsigned long flags;
 	struct ata_device *dev;
 
@@ -1152,9 +1146,7 @@ void ata_scsi_slave_destroy(struct scsi_device *sdev)
 	}
 	spin_unlock_irqrestore(ap->lock, flags);
 
-	kfree(q->dma_drain_buffer);
-	q->dma_drain_buffer = NULL;
-	q->dma_drain_size = 0;
+	kfree(sdev->dma_drain_buf);
 }
 EXPORT_SYMBOL_GPL(ata_scsi_slave_destroy);
 
diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index 4e42acbb3f32..88cac92fc153 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -978,6 +978,14 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
 		scsi_io_completion_action(cmd, result);
 }
 
+static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
+		struct request *rq)
+{
+	return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
+	       !op_is_write(req_op(rq)) &&
+	       sdev->host->hostt->dma_need_drain(rq);
+}
+
 /*
  * Function:    scsi_init_io()
  *
@@ -991,26 +999,47 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
  */
 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
 {
+	struct scsi_device *sdev = cmd->device;
 	struct request *rq = cmd->request;
+	unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
+	struct scatterlist *last_sg = NULL;
 	blk_status_t ret;
+	bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
 	int count;
 
-	if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
+	if (WARN_ON_ONCE(!nr_segs))
 		return BLK_STS_IOERR;
 
 	/*
+	 * Make sure there is space for the drain.  The driver must adjust
+	 * max_hw_segments to be prepared for this.
+	 */
+	if (need_drain)
+		nr_segs++;
+
+	/*
 	 * If sg table allocation fails, requeue request later.
 	 */
-	if (unlikely(sg_alloc_table_chained(&cmd->sdb.table,
-			blk_rq_nr_phys_segments(rq), cmd->sdb.table.sgl,
-			SCSI_INLINE_SG_CNT)))
+	if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
+			cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
 		return BLK_STS_RESOURCE;
 
 	/*
 	 * Next, walk the list, and fill in the addresses and sizes of
 	 * each segment.
 	 */
-	count = blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl);
+	count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
+
+	if (need_drain) {
+		sg_unmark_end(last_sg);
+		last_sg = sg_next(last_sg);
+		sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
+		sg_mark_end(last_sg);
+
+		rq->extra_len += sdev->dma_drain_len;
+		count++;
+	}
+
 	BUG_ON(count > cmd->sdb.table.nents);
 	cmd->sdb.table.nents = count;
 	cmd->sdb.length = blk_rq_payload_bytes(rq);