dmaengine: pl330: Optimize scatterlist transfer

Use hardware link-list instead of soft link-list to make sure
no gap among the scatter-list transfer.

Merge multiple descs [first ... last] into the last desc,
and drop the descs which have been merged into last one.

Now we get the union one rather than a soft SGL.

Obviously, this require much more MCODE buf size, because we
merge multiple descs into a union one. Increase MCODE buf size
if needed.

e.g.

  -#define MCODE_BUFF_PER_REQ	256
  +#define MCODE_BUFF_PER_REQ	512

or parsed from DT

  arm,pl330-mcbufsz-bytes = <512>;

Otherwise, you may see the warning log on the long SGL situation.

  dma-pl330 2ab90000.dmac: pl330_submit_req: Try increasing mcbufsz (403/256)

Signed-off-by: Sugar Zhang <[email protected]>
Change-Id: Iea31fc9dba08d36570d4bdbe2c5fcb61f5fed0d4

drivers/dma/pl330.c

@@ -1592,24 +1592,26 @@ static inline int _setup_xfer_cyclic(struct pl330_dmac *pl330,
  */
 static int _setup_req(struct pl330_dmac *pl330, unsigned dry_run,
 		      struct pl330_thread *thrd, unsigned index,
-		      struct _xfer_spec *pxs)
+		      struct _xfer_spec *pxs, int off)
 {
 	struct _pl330_req *req = &thrd->req[index];
 	u8 *buf = req->mc_cpu;
-	int off = 0;
 
-	PL330_DBGMC_START(req->mc_bus);
+	PL330_DBGMC_START(req->mc_bus + off);
 
 	/* DMAMOV CCR, ccr */
 	off += _emit_MOV(dry_run, &buf[off], CCR, pxs->ccr);
 
 	if (!pxs->desc->cyclic) {
 		off += _setup_xfer(pl330, dry_run, &buf[off], pxs);
 
-		/* DMASEV peripheral/event */
-		off += _emit_SEV(dry_run, &buf[off], thrd->ev);
-		/* DMAEND */
-		off += _emit_END(dry_run, &buf[off]);
+		if (pxs->desc->last) {
+			/* DMASEV peripheral/event */
+			off += _emit_SEV(dry_run, &buf[off], thrd->ev);
+
+			/* DMAEND */
+			off += _emit_END(dry_run, &buf[off]);
+		}
 	} else {
 		off += _setup_xfer_cyclic(pl330, dry_run, &buf[off],
 					  pxs, thrd->ev);
@@ -1656,7 +1658,7 @@ static inline u32 _prepare_ccr(const struct pl330_reqcfg *rqc)
  * xfer units are done or some error occurs.
  */
 static int pl330_submit_req(struct pl330_thread *thrd,
-	struct dma_pl330_desc *desc)
+	struct dma_pl330_desc *desc, int *off)
 {
 	struct pl330_dmac *pl330 = thrd->dmac;
 	struct _xfer_spec xs;
@@ -1712,12 +1714,14 @@ static int pl330_submit_req(struct pl330_thread *thrd,
 
 	idx = thrd->req[0].desc == NULL ? 0 : 1;
 
+	dev_dbg(pl330->ddma.dev, "desc-%px id %d-%d off %03d last %d cyclic %d\n",
+		desc, thrd->id, idx, *off, desc->last, desc->cyclic);
+
 	xs.ccr = ccr;
 	xs.desc = desc;
 
 	/* First dry run to check if req is acceptable */
-	ret = _setup_req(pl330, 1, thrd, idx, &xs);
-
+	ret = _setup_req(pl330, 1, thrd, idx, &xs, *off);
 	if (ret > pl330->mcbufsz / 2) {
 		dev_info(pl330->ddma.dev, "%s:%d Try increasing mcbufsz (%i/%i)\n",
 				__func__, __LINE__, ret, pl330->mcbufsz / 2);
@@ -1726,13 +1730,26 @@ static int pl330_submit_req(struct pl330_thread *thrd,
 	}
 
 	/* Hook the request */
-	thrd->lstenq = idx;
-	thrd->req[idx].desc = desc;
-	_setup_req(pl330, 0, thrd, idx, &xs);
+	if (desc->last) {
+		thrd->lstenq = idx;
+		thrd->req[idx].desc = desc;
+	}
+
+	*off = _setup_req(pl330, 0, thrd, idx, &xs, *off);
+
+	if (!desc->last) {
+		desc->status = FREE;
+		list_move_tail(&desc->node, &pl330->desc_pool);
+
+		dev_dbg(pl330->ddma.dev, "desc-%px has been merged, drop it\n", desc);
+	}
 
 	ret = 0;
 
 xfer_exit:
+	if (desc->last)
+		*off = 0;
+
 	spin_unlock_irqrestore(&pl330->lock, flags);
 
 	return ret;
@@ -2227,16 +2244,16 @@ to_desc(struct dma_async_tx_descriptor *tx)
 
 static inline void fill_queue(struct dma_pl330_chan *pch)
 {
-	struct dma_pl330_desc *desc;
-	int ret;
+	struct dma_pl330_desc *desc, *_dt;
+	int ret, off = 0;
 
-	list_for_each_entry(desc, &pch->work_list, node) {
+	list_for_each_entry_safe(desc, _dt, &pch->work_list, node) {
 
 		/* If already submitted */
 		if (desc->status == BUSY || desc->status == PAUSED)
 			continue;
 
-		ret = pl330_submit_req(pch->thread, desc);
+		ret = pl330_submit_req(pch->thread, desc, &off);
 		if (!ret) {
 			desc->status = BUSY;
 		} else if (ret == -EAGAIN) {