From aa7bde1a8b49391d34f17905a04c3acf7770283d Mon Sep 17 00:00:00 2001 From: chris hyser Date: Wed, 28 Sep 2016 12:19:50 -0700 Subject: sparc64: Enable setting "relaxed ordering" in IOMMU mappings Enable relaxed ordering for memory writes in IOMMU TSB entry from dma_4v_alloc_coherent(), dma_4v_map_page() and dma_4v_map_sg() when dma_attrs DMA_ATTR_WEAK_ORDERING is set. This requires PCI IOMMU I/O Translation Services version 2.0 API. Many PCIe devices allow enabling relaxed-ordering (memory writes bypassing other memory writes) for various DMA buffers. A notable exception is the Mellanox mlx4 IB adapter. Due to the nature of x86 HW this appears to have little performance impact there. On SPARC HW however, this results in major performance degradation getting only about 3Gbps. Enabling RO in the IOMMU entries corresponding to mlx4 data buffers increases the throughput to about 13 Gbps. Orabug: 19245907 Signed-off-by: Chris Hyser Signed-off-by: David S. Miller --- arch/sparc/include/asm/hypervisor.h | 1 + arch/sparc/kernel/pci_sun4v.c | 16 +++++++++++++++- 2 files changed, 16 insertions(+), 1 deletion(-) 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 c5c819daf800..db57d8acdc01 100644 --- a/arch/sparc/kernel/pci_sun4v.c +++ b/arch/sparc/kernel/pci_sun4v.c @@ -78,6 +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; @@ -144,6 +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; @@ -157,6 +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)) @@ -180,7 +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); @@ -277,6 +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); @@ -355,6 +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; -- cgit v1.2.3-59-g8ed1b