[PATCH 1/3 for-next] pvscsi: Use coherent memory instead of dma mapping sg lists

[Jim Gill <jgill@xxxxxxxxxx>]

Use coherent memory instead of dma mapping sg lists each
time they are used. This becomes important with SEV/swiotlb where
dma mapping otherwise implies bouncing of the data. It also gets rid
of a point of potential failure.

Tested using a "bonnie++" run on an 8GB pvscsi disk on a swiotlb=force
booted kernel.

Signed-off-by: Thomas Hellstrom <thellstrom@xxxxxxxxxx>
[jgill@xxxxxxxxxx: forwarding patch on behalf of thellstrom]
Acked-by: jgill@xxxxxxxxxx
---
 drivers/scsi/vmw_pvscsi.c | 48 +++++++++++++++++++++++------------------------
 1 file changed, 23 insertions(+), 25 deletions(-)

diff --git a/drivers/scsi/vmw_pvscsi.c b/drivers/scsi/vmw_pvscsi.c
index 8dbb4db..0573e94 100644
--- a/drivers/scsi/vmw_pvscsi.c
+++ b/drivers/scsi/vmw_pvscsi.c
@@ -27,6 +27,7 @@
 #include <linux/slab.h>
 #include <linux/workqueue.h>
 #include <linux/pci.h>
+#include <linux/dmapool.h>
 
 #include <scsi/scsi.h>
 #include <scsi/scsi_host.h>
@@ -98,6 +99,8 @@ struct pvscsi_adapter {
 
 	struct list_head		cmd_pool;
 	struct pvscsi_ctx		*cmd_map;
+
+	struct dma_pool                 *sg_pool;
 };
 
 
@@ -372,15 +375,6 @@ static int pvscsi_map_buffers(struct pvscsi_adapter *adapter,
 			pvscsi_create_sg(ctx, sg, segs);
 
 			e->flags |= PVSCSI_FLAG_CMD_WITH_SG_LIST;
-			ctx->sglPA = dma_map_single(&adapter->dev->dev,
-					ctx->sgl, SGL_SIZE, DMA_TO_DEVICE);
-			if (dma_mapping_error(&adapter->dev->dev, ctx->sglPA)) {
-				scmd_printk(KERN_ERR, cmd,
-					    "vmw_pvscsi: Failed to map ctx sglist for DMA.\n");
-				scsi_dma_unmap(cmd);
-				ctx->sglPA = 0;
-				return -ENOMEM;
-			}
 			e->dataAddr = ctx->sglPA;
 		} else
 			e->dataAddr = sg_dma_address(sg);
@@ -425,14 +419,9 @@ static void pvscsi_unmap_buffers(const struct pvscsi_adapter *adapter,
 	if (bufflen != 0) {
 		unsigned count = scsi_sg_count(cmd);
 
-		if (count != 0) {
+		if (count != 0)
 			scsi_dma_unmap(cmd);
-			if (ctx->sglPA) {
-				dma_unmap_single(&adapter->dev->dev, ctx->sglPA,
-						 SGL_SIZE, DMA_TO_DEVICE);
-				ctx->sglPA = 0;
-			}
-		} else
+		else
 			dma_unmap_single(&adapter->dev->dev, ctx->dataPA,
 					 bufflen, cmd->sc_data_direction);
 	}
@@ -1206,7 +1195,9 @@ static void pvscsi_free_sgls(const struct pvscsi_adapter *adapter)
 	unsigned i;
 
 	for (i = 0; i < adapter->req_depth; ++i, ++ctx)
-		free_pages((unsigned long)ctx->sgl, get_order(SGL_SIZE));
+		dma_pool_free(adapter->sg_pool, ctx->sgl, ctx->sglPA);
+
+	dma_pool_destroy(adapter->sg_pool);
 }
 
 static void pvscsi_shutdown_intr(struct pvscsi_adapter *adapter)
@@ -1225,10 +1216,11 @@ static void pvscsi_release_resources(struct pvscsi_adapter *adapter)
 
 	pci_release_regions(adapter->dev);
 
-	if (adapter->cmd_map) {
+	if (adapter->sg_pool)
 		pvscsi_free_sgls(adapter);
+
+	if (adapter->cmd_map)
 		kfree(adapter->cmd_map);
-	}
 
 	if (adapter->rings_state)
 		dma_free_coherent(&adapter->dev->dev, PAGE_SIZE,
@@ -1268,20 +1260,26 @@ static int pvscsi_allocate_sg(struct pvscsi_adapter *adapter)
 	struct pvscsi_ctx *ctx;
 	int i;
 
+	/* Use a dma pool so that we can impose alignment constraints. */
+	adapter->sg_pool = dma_pool_create("pvscsi_sg", pvscsi_dev(adapter),
+					   SGL_SIZE, PAGE_SIZE, 0);
+	if (!adapter->sg_pool)
+		return -ENOMEM;
+
 	ctx = adapter->cmd_map;
 	BUILD_BUG_ON(sizeof(struct pvscsi_sg_list) > SGL_SIZE);
 
 	for (i = 0; i < adapter->req_depth; ++i, ++ctx) {
-		ctx->sgl = (void *)__get_free_pages(GFP_KERNEL,
-						    get_order(SGL_SIZE));
-		ctx->sglPA = 0;
-		BUG_ON(!IS_ALIGNED(((unsigned long)ctx->sgl), PAGE_SIZE));
+		ctx->sgl = dma_pool_alloc(adapter->sg_pool, GFP_KERNEL,
+					  &ctx->sglPA);
 		if (!ctx->sgl) {
 			for (; i >= 0; --i, --ctx) {
-				free_pages((unsigned long)ctx->sgl,
-					   get_order(SGL_SIZE));
+				dma_pool_free(adapter->sg_pool, ctx->sgl,
+					      ctx->sglPA);
 				ctx->sgl = NULL;
 			}
+			dma_pool_destroy(adapter->sg_pool);
+			adapter->sg_pool = NULL;
 			return -ENOMEM;
 		}
 	}
-- 
2.7.4