diff mbox series

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

Message ID 20200812205404.GA17846@petr-dev3.eng.vmware.com (mailing list archive)
State Changes Requested
Headers show
Series [1/3,for-next] pvscsi: Use coherent memory instead of dma mapping sg lists | expand

Commit Message

Jim Gill Aug. 12, 2020, 8:54 p.m. UTC 
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@vmware.com>
[jgill@vmware.com: forwarding patch on behalf of thellstrom]
Acked-by: jgill@vmware.com
---
 drivers/scsi/vmw_pvscsi.c | 48 +++++++++++++++++++++++------------------------
 1 file changed, 23 insertions(+), 25 deletions(-)

Comments

James Bottomley Aug. 12, 2020, 8:59 p.m. UTC | #1 
This is not the correct format for sending on behalf of someone else.

First there needs to be a from at the beginning separated by a blank
line identifying the Author:

From: Thomas Hellstrom <thellstrom@vmware.com>

On Wed, 2020-08-12 at 13:54 -0700, Jim Gill wrote:
> 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@vmware.com>
> [jgill@vmware.com: forwarding patch on behalf of thellstrom]

Then we don't need this because the From: at the beginning both
preserves the author information and makes it obvious

> Acked-by: jgill@vmware.com

And finally this needs to be a Signed-off-by: to comply with the DCO
section (c).

James
Christoph Hellwig Aug. 13, 2020, 7:28 a.m. UTC | #2 
On Wed, Aug 12, 2020 at 01:54:04PM -0700, Jim Gill wrote:
> 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.

This is the wrong way around.  allocations from the coherent pool put
the system under a pointless constrained on architectures that aren't
dma coherent.  Please don't do that.
diff mbox series

Patch

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;
 		}
 	}