Merge branch 'sparc-pcie-perf'

Chris Hyser says:

====================
sparc64: PCIe performance enhancements

Ver 2 is redone for 4.8 where commit 00085f1efa387a8ce100e3734920f7639c80caa3
changed DMA attributes from struct pointer to unsigned long.

This set of patches initiates a series of PCIe performance enhancement patch
submittals.

Patch 1/2 enables version 2 of the SPARC sun4v IOMMU I/O address translation
services need for subsequent enhancements.

Patch 2/2 allows drivers to specify DMA_ATTR_WEAK_ORDERING via DMA attributes
to the SPARC DMA mapping routines enabling "relaxed ordering" for the buffer
being mapped.

[Still relevant write-up]
PCI-Express Relaxed Ordering and the Sun SPARC Enterprise M-class Servers
https://blogs.oracle.com/olympus/entry/relaxed_ordering
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

2 files changed, 42 insertions, 11 deletions

diff --git a/arch/sparc/include/asm/hypervisor.h b/arch/sparc/include/asm/hypervisor.h
index f5b6537306f0..666d5ba230d2 100644
--- a/arch/sparc/include/asm/hypervisor.h
+++ b/arch/sparc/include/asm/hypervisor.h
@@ -1744,6 +1744,7 @@ unsigned long sun4v_vintr_set_target(unsigned long dev_handle,
 
 #define HV_PCI_MAP_ATTR_READ		0x01
 #define HV_PCI_MAP_ATTR_WRITE		0x02
+#define HV_PCI_MAP_ATTR_RELAXED_ORDER	0x04
 
 #define HV_PCI_DEVICE_BUILD(b,d,f)	\
 	((((b) & 0xff) << 16) | \
diff --git a/arch/sparc/kernel/pci_sun4v.c b/arch/sparc/kernel/pci_sun4v.c
index 61c6f935accc..db57d8acdc01 100644
--- a/arch/sparc/kernel/pci_sun4v.c
+++ b/arch/sparc/kernel/pci_sun4v.c
@@ -30,8 +30,19 @@
 #define DRIVER_NAME	"pci_sun4v"
 #define PFX		DRIVER_NAME ": "
 
-static unsigned long vpci_major = 1;
-static unsigned long vpci_minor = 1;
+static unsigned long vpci_major;
+static unsigned long vpci_minor;
+
+struct vpci_version {
+	unsigned long major;
+	unsigned long minor;
+};
+
+/* Ordered from largest major to lowest */
+static struct vpci_version vpci_versions[] = {
+	{ .major = 2, .minor = 0 },
+	{ .major = 1, .minor = 1 },
+};
 
 #define PGLIST_NENTS	(PAGE_SIZE / sizeof(u64))
 
@@ -67,6 +78,10 @@ static long iommu_batch_flush(struct iommu_batch *p)
 	u64 *pglist = p->pglist;
 	unsigned long npages = p->npages;
 
+	/* VPCI maj=1, min=[0,1] only supports read and write */
+	if (vpci_major < 2)
+		prot &= (HV_PCI_MAP_ATTR_READ | HV_PCI_MAP_ATTR_WRITE);
+
 	while (npages != 0) {
 		long num;
 
@@ -133,6 +148,7 @@ static void *dma_4v_alloc_coherent(struct device *dev, size_t size,
 				   unsigned long attrs)
 {
 	unsigned long flags, order, first_page, npages, n;
+	unsigned long prot = 0;
 	struct iommu *iommu;
 	struct page *page;
 	void *ret;
@@ -146,6 +162,9 @@ static void *dma_4v_alloc_coherent(struct device *dev, size_t size,
 
 	npages = size >> IO_PAGE_SHIFT;
 
+	if (attrs & DMA_ATTR_WEAK_ORDERING)
+		prot = HV_PCI_MAP_ATTR_RELAXED_ORDER;
+
 	nid = dev->archdata.numa_node;
 	page = alloc_pages_node(nid, gfp, order);
 	if (unlikely(!page))
@@ -169,7 +188,7 @@ static void *dma_4v_alloc_coherent(struct device *dev, size_t size,
 	local_irq_save(flags);
 
 	iommu_batch_start(dev,
-			  (HV_PCI_MAP_ATTR_READ |
+			  (HV_PCI_MAP_ATTR_READ | prot |
 			   HV_PCI_MAP_ATTR_WRITE),
 			  entry);
 
@@ -266,6 +285,9 @@ static dma_addr_t dma_4v_map_page(struct device *dev, struct page *page,
 	if (direction != DMA_TO_DEVICE)
 		prot |= HV_PCI_MAP_ATTR_WRITE;
 
+	if (attrs & DMA_ATTR_WEAK_ORDERING)
+		prot |= HV_PCI_MAP_ATTR_RELAXED_ORDER;
+
 	local_irq_save(flags);
 
 	iommu_batch_start(dev, prot, entry);
@@ -344,6 +366,9 @@ static int dma_4v_map_sg(struct device *dev, struct scatterlist *sglist,
 	if (direction != DMA_TO_DEVICE)
 		prot |= HV_PCI_MAP_ATTR_WRITE;
 
+	if (attrs & DMA_ATTR_WEAK_ORDERING)
+		prot |= HV_PCI_MAP_ATTR_RELAXED_ORDER;
+
 	outs = s = segstart = &sglist[0];
 	outcount = 1;
 	incount = nelems;
@@ -907,22 +932,27 @@ static int pci_sun4v_probe(struct platform_device *op)
 	struct device_node *dp;
 	struct iommu *iommu;
 	u32 devhandle;
-	int i, err;
+	int i, err = -ENODEV;
 
 	dp = op->dev.of_node;
 
 	if (!hvapi_negotiated++) {
-		err = sun4v_hvapi_register(HV_GRP_PCI,
-					   vpci_major,
-					   &vpci_minor);
+		for (i = 0; i < ARRAY_SIZE(vpci_versions); i++) {
+			vpci_major = vpci_versions[i].major;
+			vpci_minor = vpci_versions[i].minor;
+
+			err = sun4v_hvapi_register(HV_GRP_PCI, vpci_major,
+						   &vpci_minor);
+			if (!err)
+				break;
+		}
 
 		if (err) {
-			printk(KERN_ERR PFX "Could not register hvapi, "
-			       "err=%d\n", err);
+			pr_err(PFX "Could not register hvapi, err=%d\n", err);
 			return err;
 		}
-		printk(KERN_INFO PFX "Registered hvapi major[%lu] minor[%lu]\n",
-		       vpci_major, vpci_minor);
+		pr_info(PFX "Registered hvapi major[%lu] minor[%lu]\n",
+			vpci_major, vpci_minor);
 
 		dma_ops = &sun4v_dma_ops;
 	}