[PATCH 0/2 v2] mtip32xx: Fix arch-specific performance issues

[PATCH 0/2 v2] mtip32xx: Fix arch-specific performance issues

[Felipe Franciosi]

This V2 series reintroduces blk_queue_bounce(), but also sets a correct
blk_queue_bounce_limit() during the device initialisation. I have tested
this series and confirmed that performance is good on a 32-bit dom0 on Xen.

[PATCH 1/2] mtip32xx: Set queue bounce limit

[Felipe Franciosi]

We need to set the queue bounce limit during the device initialization to
prevent excessive bouncing on 32 bit architectures.

Signed-off-by: Felipe Franciosi <felipe@paradoxo.org>
---
 drivers/block/mtip32xx/mtip32xx.c |    1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/block/mtip32xx/mtip32xx.c b/drivers/block/mtip32xx/mtip32xx.c
index 5160269..787c9d3 100644
--- a/drivers/block/mtip32xx/mtip32xx.c
+++ b/drivers/block/mtip32xx/mtip32xx.c
@@ -4213,6 +4213,7 @@ skip_create_disk:
 	blk_queue_max_hw_sectors(dd->queue, 0xffff);
 	blk_queue_max_segment_size(dd->queue, 0x400000);
 	blk_queue_io_min(dd->queue, 4096);
+	blk_queue_bounce_limit(dd->queue, dd->pdev->dma_mask);
 
 	/*
 	 * write back cache is not supported in the device. FUA depends on
-- 
1.7.10.4

[PATCH 2/2] mtip32xx: Unmap the DMA segments before completing the IO request

[Felipe Franciosi]

If the buffers are unmapped after completing a request, then stale data
might be in the request.

Signed-off-by: Felipe Franciosi <felipe@paradoxo.org>
---
 drivers/block/mtip32xx/mtip32xx.c |   24 ++++++++++++------------
 1 file changed, 12 insertions(+), 12 deletions(-)

diff --git a/drivers/block/mtip32xx/mtip32xx.c b/drivers/block/mtip32xx/mtip32xx.c
index 787c9d3..4dd2642 100644
--- a/drivers/block/mtip32xx/mtip32xx.c
+++ b/drivers/block/mtip32xx/mtip32xx.c
@@ -266,6 +266,12 @@ static void mtip_async_complete(struct mtip_port *port,
 			"Command tag %d failed due to TFE\n", tag);
 	}
 
+	/* Unmap the DMA scatter list entries */
+	dma_unmap_sg(&dd->pdev->dev,
+		command->sg,
+		command->scatter_ents,
+		command->direction);
+
 	/* Upper layer callback */
 	if (likely(command->async_callback))
 		command->async_callback(command->async_data, cb_status);
@@ -273,12 +279,6 @@ static void mtip_async_complete(struct mtip_port *port,
 	command->async_callback = NULL;
 	command->comp_func = NULL;
 
-	/* Unmap the DMA scatter list entries */
-	dma_unmap_sg(&dd->pdev->dev,
-		command->sg,
-		command->scatter_ents,
-		command->direction);
-
 	/* Clear the allocated and active bits for the command */
 	atomic_set(&port->commands[tag].active, 0);
 	release_slot(port, tag);
@@ -709,6 +709,12 @@ static void mtip_timeout_function(unsigned long int data)
 			 */
 			writel(1 << bit, port->completed[group]);
 
+			/* Unmap the DMA scatter list entries */
+			dma_unmap_sg(&port->dd->pdev->dev,
+					command->sg,
+					command->scatter_ents,
+					command->direction);
+
 			/* Call the async completion callback. */
 			if (likely(command->async_callback))
 				command->async_callback(command->async_data,
@@ -716,12 +722,6 @@ static void mtip_timeout_function(unsigned long int data)
 			command->async_callback = NULL;
 			command->comp_func = NULL;
 
-			/* Unmap the DMA scatter list entries */
-			dma_unmap_sg(&port->dd->pdev->dev,
-					command->sg,
-					command->scatter_ents,
-					command->direction);
-
 			/*
 			 * Clear the allocated bit and active tag for the
 			 * command.
-- 
1.7.10.4

Re: [PATCH 2/2] mtip32xx: Unmap the DMA segments before completing the IO request

[Jens Axboe]

On 03/13/2014 08:34 AM, Felipe Franciosi wrote:
> If the buffers are unmapped after completing a request, then stale data
> might be in the request.

Both should be marked for stable, I'll add that.

Sam, I'll queue these two up. Please re-send your fix for bug introduced 
with the unaligned reduction as a proper patch, and I'll get that in too.

-- 
Jens Axboe