usb: musb: ux500: optimize DMA callback routine

Skip the use of work queue and call musb_dma_completion() directly from
DMA callback context.

Here follows measurements on a Snowball board with ondemand governor active.

Performance using work queue:
(105 MB) copied, 6.23758 s, 16.8 MB/s
(105 MB) copied, 5.7151 s, 18.3 MB/s
(105 MB) copied, 5.83583 s, 18.0 MB/s
(105 MB) copied, 5.93611 s, 17.7 MB/s

Performance without work queue
(105 MB) copied, 5.62173 s, 18.7 MB/s
(105 MB) copied, 5.61811 s, 18.7 MB/s
(105 MB) copied, 5.57817 s, 18.8 MB/s
(105 MB) copied, 5.58549 s, 18.8 MB/s
Signed-off-by: Per Forlin <per.forlin@linaro.org>
Acked-by: Mian Yousaf Kaukab <mian.yousaf.kaukab@stericsson.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>

drivers/usb/musb/ux500_dma.c

@@ -37,7 +37,6 @@ struct ux500_dma_channel {
 	struct dma_channel channel;
 	struct ux500_dma_controller *controller;
 	struct musb_hw_ep *hw_ep;
-	struct work_struct channel_work;
 	struct dma_chan *dma_chan;
 	unsigned int cur_len;
 	dma_cookie_t cookie;
@@ -56,31 +55,11 @@ struct ux500_dma_controller {
 	dma_addr_t phy_base;
 };
 
-/* Work function invoked from DMA callback to handle tx transfers. */
-static void ux500_tx_work(struct work_struct *data)
-{
-	struct ux500_dma_channel *ux500_channel = container_of(data,
-		struct ux500_dma_channel, channel_work);
-	struct musb_hw_ep       *hw_ep = ux500_channel->hw_ep;
-	struct musb *musb = hw_ep->musb;
-	unsigned long flags;
-
-	dev_dbg(musb->controller, "DMA tx transfer done on hw_ep=%d\n",
-		hw_ep->epnum);
-
-	spin_lock_irqsave(&musb->lock, flags);
-	ux500_channel->channel.actual_len = ux500_channel->cur_len;
-	ux500_channel->channel.status = MUSB_DMA_STATUS_FREE;
-	musb_dma_completion(musb, hw_ep->epnum,
-				ux500_channel->is_tx);
-	spin_unlock_irqrestore(&musb->lock, flags);
-}
-
 /* Work function invoked from DMA callback to handle rx transfers. */
-static void ux500_rx_work(struct work_struct *data)
+void ux500_dma_callback(void *private_data)
 {
-	struct ux500_dma_channel *ux500_channel = container_of(data,
-		struct ux500_dma_channel, channel_work);
+	struct dma_channel *channel = private_data;
+	struct ux500_dma_channel *ux500_channel = channel->private_data;
 	struct musb_hw_ep       *hw_ep = ux500_channel->hw_ep;
 	struct musb *musb = hw_ep->musb;
 	unsigned long flags;
@@ -94,14 +73,7 @@ static void ux500_rx_work(struct work_struct *data)
 	musb_dma_completion(musb, hw_ep->epnum,
 		ux500_channel->is_tx);
 	spin_unlock_irqrestore(&musb->lock, flags);
-}
-
-void ux500_dma_callback(void *private_data)
-{
-	struct dma_channel *channel = (struct dma_channel *)private_data;
-	struct ux500_dma_channel *ux500_channel = channel->private_data;
 
-	schedule_work(&ux500_channel->channel_work);
 }
 
 static bool ux500_configure_channel(struct dma_channel *channel,
@@ -330,7 +302,6 @@ static int ux500_dma_controller_start(struct dma_controller *c)
 	void **param_array;
 	struct ux500_dma_channel *channel_array;
 	u32 ch_count;
-	void (*musb_channel_work)(struct work_struct *);
 	dma_cap_mask_t mask;
 
 	if ((data->num_rx_channels > UX500_MUSB_DMA_NUM_RX_CHANNELS) ||
@@ -347,7 +318,6 @@ static int ux500_dma_controller_start(struct dma_controller *c)
 	channel_array = controller->rx_channel;
 	ch_count = data->num_rx_channels;
 	param_array = data->dma_rx_param_array;
-	musb_channel_work = ux500_rx_work;
 
 	for (dir = 0; dir < 2; dir++) {
 		for (ch_num = 0; ch_num < ch_count; ch_num++) {
@@ -374,15 +344,12 @@ static int ux500_dma_controller_start(struct dma_controller *c)
 				return -EBUSY;
 			}
 
-			INIT_WORK(&ux500_channel->channel_work,
-				musb_channel_work);
 		}
 
 		/* Prepare the loop for TX channels */
 		channel_array = controller->tx_channel;
 		ch_count = data->num_tx_channels;
 		param_array = data->dma_tx_param_array;
-		musb_channel_work = ux500_tx_work;
 		is_tx = 1;
 	}