[PATCH 4/5] switchtec: Improve MRPC efficiency by leveraging write combining

From: Wesley Sheng <wesley.sheng@microchip.com>
To: <kurt.schwemmer@microsemi.com>, <logang@deltatee.com>,
	<bhelgaas@google.com>, <linux-pci@vger.kernel.org>,
	<linux-kernel@vger.kernel.org>
Cc: <wesleyshenggit@sina.com>, <wesley.sheng@microchip.com>
Subject: [PATCH 4/5] switchtec: Improve MRPC efficiency by leveraging write combining
Date: Mon, 10 Dec 2018 17:12:23 +0800	[thread overview]
Message-ID: <1544433144-7563-5-git-send-email-wesley.sheng@microchip.com> (raw)
In-Reply-To: <1544433144-7563-1-git-send-email-wesley.sheng@microchip.com>

From: Kelvin Cao <kelvin.cao@microchip.com>

MRPC Input buffer is mostly memory without any side effects, thus we can
improve the access time by enabling write combining on only this region of
the BAR.

In a few places, we still need to flush the WC buffer. To do this, we
simply read from the Outbound Doorbell register seeing reads to this
register are processed by low latency hardware.

Signed-off-by: Kelvin Cao <kelvin.cao@microchip.com>
Signed-off-by: Wesley Sheng <wesley.sheng@microchip.com>
Reviewed-by: Logan Gunthorpe <logang@deltatee.com>
---
 drivers/pci/switch/switchtec.c | 41 +++++++++++++++++++++++++++++++++++------
 1 file changed, 35 insertions(+), 6 deletions(-)

diff --git a/drivers/pci/switch/switchtec.c b/drivers/pci/switch/switchtec.c
index a908670..0b8862b 100644
--- a/drivers/pci/switch/switchtec.c
+++ b/drivers/pci/switch/switchtec.c
@@ -113,6 +113,19 @@ static void stuser_set_state(struct switchtec_user *stuser,
 
 static void mrpc_complete_cmd(struct switchtec_dev *stdev);
 
+static void flush_wc_buf(struct switchtec_dev *stdev)
+{
+	struct ntb_dbmsg_regs __iomem *mmio_dbmsg;
+
+	/*
+	 * odb (outbound doorbell) register is processed by low latency
+	 * hardware and w/o side effect
+	 */
+	mmio_dbmsg = (void __iomem *)stdev->mmio_ntb +
+		SWITCHTEC_NTB_REG_DBMSG_OFFSET;
+	ioread32(&mmio_dbmsg->odb);
+}
+
 static void mrpc_cmd_submit(struct switchtec_dev *stdev)
 {
 	/* requires the mrpc_mutex to already be held when called */
@@ -132,6 +145,7 @@ static void mrpc_cmd_submit(struct switchtec_dev *stdev)
 	stdev->mrpc_busy = 1;
 	memcpy_toio(&stdev->mmio_mrpc->input_data,
 		    stuser->data, stuser->data_len);
+	flush_wc_buf(stdev);
 	iowrite32(stuser->cmd, &stdev->mmio_mrpc->cmd);
 
 	schedule_delayed_work(&stdev->mrpc_timeout,
@@ -1231,23 +1245,38 @@ static int switchtec_init_pci(struct switchtec_dev *stdev,
 			      struct pci_dev *pdev)
 {
 	int rc;
+	void __iomem *map;
+	unsigned long res_start, res_len;
 
 	rc = pcim_enable_device(pdev);
 	if (rc)
 		return rc;
 
-	rc = pcim_iomap_regions(pdev, 0x1, KBUILD_MODNAME);
-	if (rc)
-		return rc;
-
 	rc = dma_set_coherent_mask(&pdev->dev, DMA_BIT_MASK(64));
 	if (rc)
 		return rc;
 
 	pci_set_master(pdev);
 
-	stdev->mmio = pcim_iomap_table(pdev)[0];
-	stdev->mmio_mrpc = stdev->mmio + SWITCHTEC_GAS_MRPC_OFFSET;
+	res_start = pci_resource_start(pdev, 0);
+	res_len = pci_resource_len(pdev, 0);
+
+	if (!devm_request_mem_region(&pdev->dev, res_start,
+				     res_len, KBUILD_MODNAME))
+		return -EBUSY;
+
+	stdev->mmio_mrpc = devm_ioremap_wc(&pdev->dev, res_start,
+					   SWITCHTEC_GAS_TOP_CFG_OFFSET);
+	if (!stdev->mmio_mrpc)
+		return -ENOMEM;
+
+	map = devm_ioremap(&pdev->dev,
+			   res_start + SWITCHTEC_GAS_TOP_CFG_OFFSET,
+			   res_len - SWITCHTEC_GAS_TOP_CFG_OFFSET);
+	if (!map)
+		return -ENOMEM;
+
+	stdev->mmio = map - SWITCHTEC_GAS_TOP_CFG_OFFSET;
 	stdev->mmio_sw_event = stdev->mmio + SWITCHTEC_GAS_SW_EVENT_OFFSET;
 	stdev->mmio_sys_info = stdev->mmio + SWITCHTEC_GAS_SYS_INFO_OFFSET;
 	stdev->mmio_flash_info = stdev->mmio + SWITCHTEC_GAS_FLASH_INFO_OFFSET;
-- 
2.7.4

