Ported pcie-brcmstb bounce buffer implementation to ARM64. (#3144)

Ported pcie-brcmstb bounce buffer implementation to ARM64.
This enables full 4G RAM usage on Raspberry Pi in 64-bit mode.

Signed-off-by: Yaroslav Rosomakho <yaroslavros@gmail.com>

arch/arm64/mm/dma-mapping.c

@@ -160,37 +160,6 @@ void arch_sync_dma_for_cpu(struct device *dev, phys_addr_t paddr,
 	__dma_unmap_area(phys_to_virt(paddr), size, dir);
 }
 
-#ifdef CONFIG_IOMMU_DMA
-static int __swiotlb_get_sgtable_page(struct sg_table *sgt,
-				      struct page *page, size_t size)
-{
-	int ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
-
-	if (!ret)
-		sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
-
-	return ret;
-}
-
-static int __swiotlb_mmap_pfn(struct vm_area_struct *vma,
-			      unsigned long pfn, size_t size)
-{
-	int ret = -ENXIO;
-	unsigned long nr_vma_pages = vma_pages(vma);
-	unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
-	unsigned long off = vma->vm_pgoff;
-
-	if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) {
-		ret = remap_pfn_range(vma, vma->vm_start,
-				      pfn + off,
-				      vma->vm_end - vma->vm_start,
-				      vma->vm_page_prot);
-	}
-
-	return ret;
-}
-#endif /* CONFIG_IOMMU_DMA */
-
 static int __init atomic_pool_init(void)
 {
 	pgprot_t prot = __pgprot(PROT_NORMAL_NC);
@@ -579,9 +548,9 @@ static dma_addr_t __iommu_map_page(struct device *dev, struct page *page,
 	int prot = dma_info_to_prot(dir, coherent, attrs);
 	dma_addr_t dev_addr = iommu_dma_map_page(dev, page, offset, size, prot);
 
-	if (!coherent && !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
+	if (!iommu_dma_mapping_error(dev, dev_addr) &&
-	    !iommu_dma_mapping_error(dev, dev_addr))
+	    (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
-		__dma_map_area(page_address(page) + offset, size, dir);
+		__iommu_sync_single_for_device(dev, dev_addr, size, dir);
 
 	return dev_addr;
 }

drivers/pci/controller/Makefile

@@ -32,6 +32,9 @@ obj-$(CONFIG_PCIE_BRCMSTB) += pcie-brcmstb.o
 ifdef CONFIG_ARM
 obj-$(CONFIG_PCIE_BRCMSTB) += pcie-brcmstb-bounce.o
 endif
+ifdef CONFIG_ARM64
+obj-$(CONFIG_PCIE_BRCMSTB) += pcie-brcmstb-bounce64.o
+endif
 
 obj-$(CONFIG_VMD) += vmd.o
 # pcie-hisi.o quirks are needed even without CONFIG_PCIE_DW

drivers/pci/controller/pcie-brcmstb-bounce.h

@@ -6,7 +6,7 @@
 #ifndef _PCIE_BRCMSTB_BOUNCE_H
 #define _PCIE_BRCMSTB_BOUNCE_H
 
-#ifdef CONFIG_ARM
+#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
 
 int brcm_pcie_bounce_init(struct device *dev, unsigned long buffer_size,
 			  dma_addr_t threshold);

drivers/pci/controller/pcie-brcmstb-bounce64.c

@@ -0,0 +1,569 @@
+/*
+ *  This code started out as a version of arch/arm/common/dmabounce.c,
+ *  modified to cope with highmem pages. Now it has been changed heavily -
+ *  it now preallocates a large block (currently 4MB) and carves it up
+ *  sequentially in ring fashion, and DMA is used to copy the data - to the
+ *  point where very little of the original remains.
+ *
+ *  Copyright (C) 2019 Raspberry Pi (Trading) Ltd.
+ *
+ *  Original version by Brad Parker (brad@heeltoe.com)
+ *  Re-written by Christopher Hoover <ch@murgatroid.com>
+ *  Made generic by Deepak Saxena <dsaxena@plexity.net>
+ *
+ *  Copyright (C) 2002 Hewlett Packard Company.
+ *  Copyright (C) 2004 MontaVista Software, Inc.
+ *
+ *  This program is free software; you can redistribute it and/or
+ *  modify it under the terms of the GNU General Public License
+ *  version 2 as published by the Free Software Foundation.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/slab.h>
+#include <linux/page-flags.h>
+#include <linux/device.h>
+#include <linux/dma-mapping.h>
+#include <linux/dma-direct.h>
+#include <linux/dmapool.h>
+#include <linux/list.h>
+#include <linux/scatterlist.h>
+#include <linux/bitmap.h>
+#include <linux/swiotlb.h>
+
+#include <asm/cacheflush.h>
+
+#define STATS
+
+#ifdef STATS
+#define DO_STATS(X) do { X ; } while (0)
+#else
+#define DO_STATS(X) do { } while (0)
+#endif
+
+/* ************************************************** */
+
+struct safe_buffer {
+	struct list_head node;
+
+	/* original request */
+	size_t		size;
+	int		direction;
+
+	struct dmabounce_pool *pool;
+	void		*safe;
+	dma_addr_t	unsafe_dma_addr;
+	dma_addr_t	safe_dma_addr;
+};
+
+struct dmabounce_pool {
+	unsigned long	pages;
+	void		*virt_addr;
+	dma_addr_t	dma_addr;
+	unsigned long	*alloc_map;
+	unsigned long	alloc_pos;
+	spinlock_t	lock;
+	struct device	*dev;
+	unsigned long   num_pages;
+#ifdef STATS
+	size_t		max_size;
+	unsigned long	num_bufs;
+	unsigned long   max_bufs;
+	unsigned long   max_pages;
+#endif
+};
+
+struct dmabounce_device_info {
+	struct device *dev;
+	dma_addr_t threshold;
+	struct list_head safe_buffers;
+	struct dmabounce_pool pool;
+	rwlock_t lock;
+#ifdef STATS
+	unsigned long map_count;
+	unsigned long unmap_count;
+	unsigned long sync_dev_count;
+	unsigned long sync_cpu_count;
+	unsigned long fail_count;
+	int attr_res;
+#endif
+};
+
+static struct dmabounce_device_info *g_dmabounce_device_info;
+
+extern int bcm2838_dma40_memcpy_init(void);
+extern void bcm2838_dma40_memcpy(dma_addr_t dst, dma_addr_t src, size_t size);
+
+#ifdef STATS
+static ssize_t
+bounce_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct dmabounce_device_info *device_info = g_dmabounce_device_info;
+	return sprintf(buf, "m:%lu/%lu s:%lu/%lu f:%lu s:%zu b:%lu/%lu a:%lu/%lu\n",
+		device_info->map_count,
+		device_info->unmap_count,
+		device_info->sync_dev_count,
+		device_info->sync_cpu_count,
+		device_info->fail_count,
+		device_info->pool.max_size,
+		device_info->pool.num_bufs,
+		device_info->pool.max_bufs,
+		device_info->pool.num_pages * PAGE_SIZE,
+		device_info->pool.max_pages * PAGE_SIZE);
+}
+
+static DEVICE_ATTR(dmabounce_stats, 0444, bounce_show, NULL);
+#endif
+
+static int bounce_create(struct dmabounce_pool *pool, struct device *dev,
+			 unsigned long buffer_size)
+{
+	int ret = -ENOMEM;
+	pool->pages = (buffer_size + PAGE_SIZE - 1)/PAGE_SIZE;
+	pool->alloc_map = bitmap_zalloc(pool->pages, GFP_KERNEL);
+	if (!pool->alloc_map)
+		goto err_bitmap;
+	pool->virt_addr = dma_alloc_coherent(dev, pool->pages * PAGE_SIZE,
+					     &pool->dma_addr, GFP_KERNEL);
+	if (!pool->virt_addr)
+		goto err_dmabuf;
+
+	pool->alloc_pos = 0;
+	spin_lock_init(&pool->lock);
+	pool->dev = dev;
+	pool->num_pages = 0;
+
+	DO_STATS(pool->max_size = 0);
+	DO_STATS(pool->num_bufs = 0);
+	DO_STATS(pool->max_bufs = 0);
+	DO_STATS(pool->max_pages = 0);
+
+	return  0;
+
+err_dmabuf:
+	bitmap_free(pool->alloc_map);
+err_bitmap:
+	return ret;
+}
+
+static void bounce_destroy(struct dmabounce_pool *pool)
+{
+	dma_free_coherent(pool->dev, pool->pages * PAGE_SIZE, pool->virt_addr,
+			  pool->dma_addr);
+
+	bitmap_free(pool->alloc_map);
+}
+
+static void *bounce_alloc(struct dmabounce_pool *pool, size_t size,
+			  dma_addr_t *dmaaddrp)
+{
+	unsigned long pages;
+	unsigned long flags;
+	unsigned long pos;
+
+	pages = (size + PAGE_SIZE - 1)/PAGE_SIZE;
+
+	DO_STATS(pool->max_size = max(size, pool->max_size));
+
+	spin_lock_irqsave(&pool->lock, flags);
+	pos = bitmap_find_next_zero_area(pool->alloc_map, pool->pages,
+					 pool->alloc_pos, pages, 0);
+	/* If not found, try from the start */
+	if (pos >= pool->pages && pool->alloc_pos)
+		pos = bitmap_find_next_zero_area(pool->alloc_map, pool->pages,
+						 0, pages, 0);
+
+	if (pos >= pool->pages) {
+		spin_unlock_irqrestore(&pool->lock, flags);
+		return NULL;
+	}
+
+	bitmap_set(pool->alloc_map, pos, pages);
+	pool->alloc_pos = (pos + pages) % pool->pages;
+	pool->num_pages += pages;
+
+	DO_STATS(pool->num_bufs++);
+	DO_STATS(pool->max_bufs = max(pool->num_bufs, pool->max_bufs));
+	DO_STATS(pool->max_pages = max(pool->num_pages, pool->max_pages));
+
+	spin_unlock_irqrestore(&pool->lock, flags);
+
+	*dmaaddrp = pool->dma_addr + pos * PAGE_SIZE;
+
+	return pool->virt_addr + pos * PAGE_SIZE;
+}
+
+static void
+bounce_free(struct dmabounce_pool *pool, void *buf, size_t size)
+{
+	unsigned long pages;
+	unsigned long flags;
+	unsigned long pos;
+
+	pages = (size + PAGE_SIZE - 1)/PAGE_SIZE;
+	pos = (buf - pool->virt_addr)/PAGE_SIZE;
+
+	BUG_ON((buf - pool->virt_addr) & (PAGE_SIZE - 1));
+
+	spin_lock_irqsave(&pool->lock, flags);
+	bitmap_clear(pool->alloc_map, pos, pages);
+	pool->num_pages -= pages;
+	if (pool->num_pages == 0)
+		pool->alloc_pos = 0;
+	DO_STATS(pool->num_bufs--);
+	spin_unlock_irqrestore(&pool->lock, flags);
+}
+
+/* allocate a 'safe' buffer and keep track of it */
+static struct safe_buffer *
+alloc_safe_buffer(struct dmabounce_device_info *device_info,
+		  dma_addr_t dma_addr, size_t size, enum dma_data_direction dir)
+{
+	struct safe_buffer *buf;
+	struct dmabounce_pool *pool = &device_info->pool;
+	struct device *dev = device_info->dev;
+	unsigned long flags;
+
+	/*
+	 * Although one might expect this to be called in thread context,
+	 * using GFP_KERNEL here leads to hard-to-debug lockups. in_atomic()
+	 * was previously used to select the appropriate allocation mode,
+	 * but this is unsafe.
+	 */
+	buf = kmalloc(sizeof(struct safe_buffer), GFP_ATOMIC);
+	if (!buf) {
+		dev_warn(dev, "%s: kmalloc failed\n", __func__);
+		return NULL;
+	}
+
+	buf->unsafe_dma_addr = dma_addr;
+	buf->size = size;
+	buf->direction = dir;
+	buf->pool = pool;
+
+	buf->safe = bounce_alloc(pool, size, &buf->safe_dma_addr);
+
+	if (!buf->safe) {
+		dev_warn(dev,
+			 "%s: could not alloc dma memory (size=%zu)\n",
+			 __func__, size);
+		kfree(buf);
+		return NULL;
+	}
+
+	write_lock_irqsave(&device_info->lock, flags);
+	list_add(&buf->node, &device_info->safe_buffers);
+	write_unlock_irqrestore(&device_info->lock, flags);
+
+	return buf;
+}
+
+/* determine if a buffer is from our "safe" pool */
+static struct safe_buffer *
+find_safe_buffer(struct dmabounce_device_info *device_info,
+		 dma_addr_t safe_dma_addr)
+{
+	struct safe_buffer *b, *rb = NULL;
+	unsigned long flags;
+
+	read_lock_irqsave(&device_info->lock, flags);
+
+	list_for_each_entry(b, &device_info->safe_buffers, node)
+		if (b->safe_dma_addr <= safe_dma_addr &&
+		    b->safe_dma_addr + b->size > safe_dma_addr) {
+			rb = b;
+			break;
+		}
+
+	read_unlock_irqrestore(&device_info->lock, flags);
+	return rb;
+}
+
+static void
+free_safe_buffer(struct dmabounce_device_info *device_info,
+		 struct safe_buffer *buf)
+{
+	unsigned long flags;
+
+	write_lock_irqsave(&device_info->lock, flags);
+	list_del(&buf->node);
+	write_unlock_irqrestore(&device_info->lock, flags);
+
+	bounce_free(buf->pool, buf->safe, buf->size);
+
+	kfree(buf);
+}
+
+/* ************************************************** */
+
+static struct safe_buffer *
+find_safe_buffer_dev(struct device *dev, dma_addr_t dma_addr, const char *where)
+{
+	if (!dev || !g_dmabounce_device_info)
+		return NULL;
+	if (dma_mapping_error(dev, dma_addr)) {
+		dev_err(dev, "Trying to %s invalid mapping\n", where);
+		return NULL;
+	}
+	return find_safe_buffer(g_dmabounce_device_info, dma_addr);
+}
+
+static dma_addr_t
+map_single(struct device *dev, struct safe_buffer *buf, size_t size,
+	   enum dma_data_direction dir, unsigned long attrs)
+{
+	BUG_ON(buf->size != size);
+	BUG_ON(buf->direction != dir);
+
+	dev_dbg(dev, "map: %llx->%llx\n", (u64)buf->unsafe_dma_addr,
+		(u64)buf->safe_dma_addr);
+
+	if ((dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL) &&
+	    !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
+		bcm2838_dma40_memcpy(buf->safe_dma_addr, buf->unsafe_dma_addr,
+				     size);
+
+	return buf->safe_dma_addr;
+}
+
+static dma_addr_t
+unmap_single(struct device *dev, struct safe_buffer *buf, size_t size,
+	     enum dma_data_direction dir, unsigned long attrs)
+{
+	BUG_ON(buf->size != size);
+	BUG_ON(buf->direction != dir);
+
+	if ((dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL) &&
+	    !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
+		dev_dbg(dev, "unmap: %llx->%llx\n", (u64)buf->safe_dma_addr,
+			(u64)buf->unsafe_dma_addr);
+
+		bcm2838_dma40_memcpy(buf->unsafe_dma_addr, buf->safe_dma_addr,
+				     size);
+	}
+	return buf->unsafe_dma_addr;
+}
+
+/* ************************************************** */
+
+/*
+ * see if a buffer address is in an 'unsafe' range.  if it is
+ * allocate a 'safe' buffer and copy the unsafe buffer into it.
+ * substitute the safe buffer for the unsafe one.
+ * (basically move the buffer from an unsafe area to a safe one)
+ */
+static dma_addr_t
+dmabounce_map_page(struct device *dev, struct page *page, unsigned long offset,
+		   size_t size, enum dma_data_direction dir,
+		   unsigned long attrs)
+{
+	struct dmabounce_device_info *device_info = g_dmabounce_device_info;
+	dma_addr_t dma_addr;
+
+	dma_addr = phys_to_dma(dev, page_to_phys(page)) + offset;
+
+	swiotlb_sync_single_for_device(dev, dma_addr, size, dir);
+        if (!is_device_dma_coherent(dev))
+		__dma_map_area(phys_to_virt(dma_to_phys(dev, dma_addr)), size, dir);
+
+	if (device_info && (dma_addr + size) > device_info->threshold) {
+		struct safe_buffer *buf;
+
+		buf = alloc_safe_buffer(device_info, dma_addr, size, dir);
+		if (!buf) {
+			DO_STATS(device_info->fail_count++);
+			return (~(dma_addr_t)0x0);
+		}
+
+		DO_STATS(device_info->map_count++);
+
+		dma_addr = map_single(dev, buf, size, dir, attrs);
+	}
+	return dma_addr;
+}
+
+/*
+ * see if a mapped address was really a "safe" buffer and if so, copy
+ * the data from the safe buffer back to the unsafe buffer and free up
+ * the safe buffer.  (basically return things back to the way they
+ * should be)
+ */
+static void
+dmabounce_unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
+		     enum dma_data_direction dir, unsigned long attrs)
+{
+	struct safe_buffer *buf;
+
+	buf = find_safe_buffer_dev(dev, dma_addr, __func__);
+	if (buf) {
+		DO_STATS(g_dmabounce_device_info->unmap_count++);
+		dma_addr = unmap_single(dev, buf, size, dir, attrs);
+		free_safe_buffer(g_dmabounce_device_info, buf);
+	}
+
+        if (!is_device_dma_coherent(dev))
+		__dma_unmap_area(phys_to_virt(dma_to_phys(dev, dma_addr)), size, dir);
+	swiotlb_sync_single_for_cpu(dev, dma_addr, size, dir);
+}
+
+/*
+ * A version of dmabounce_map_page that assumes the mapping has already
+ * been created - intended for streaming operation.
+ */
+static void
+dmabounce_sync_for_device(struct device *dev, dma_addr_t dma_addr, size_t size,
+			  enum dma_data_direction dir)
+{
+	struct safe_buffer *buf;
+
+        swiotlb_sync_single_for_device(dev, dma_addr, size, dir);
+        if (!is_device_dma_coherent(dev))
+                __dma_map_area(phys_to_virt(dma_to_phys(dev, dma_addr)), size, dir);
+
+	buf = find_safe_buffer_dev(dev, dma_addr, __func__);
+	if (buf) {
+		DO_STATS(g_dmabounce_device_info->sync_dev_count++);
+		map_single(dev, buf, size, dir, 0);
+	}
+}
+
+/*
+ * A version of dmabounce_unmap_page that doesn't destroy the mapping -
+ * intended for streaming operation.
+ */
+static void
+dmabounce_sync_for_cpu(struct device *dev, dma_addr_t dma_addr,
+		       size_t size, enum dma_data_direction dir)
+{
+	struct safe_buffer *buf;
+
+	buf = find_safe_buffer_dev(dev, dma_addr, __func__);
+	if (buf) {
+		DO_STATS(g_dmabounce_device_info->sync_cpu_count++);
+		dma_addr = unmap_single(dev, buf, size, dir, 0);
+	}
+
+        if (!is_device_dma_coherent(dev))
+                __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dma_addr)), size, dir);
+        swiotlb_sync_single_for_cpu(dev, dma_addr, size, dir);
+}
+
+static int dmabounce_dma_supported(struct device *dev, u64 dma_mask)
+{
+	if (g_dmabounce_device_info)
+		return 0;
+
+	return swiotlb_dma_supported(dev, dma_mask);
+}
+
+static const struct dma_map_ops dmabounce_ops = {
+	.alloc			= dma_direct_alloc,
+	.free			= dma_direct_free,
+	.map_page		= dmabounce_map_page,
+	.unmap_page		= dmabounce_unmap_page,
+	.sync_single_for_cpu	= dmabounce_sync_for_cpu,
+	.sync_single_for_device	= dmabounce_sync_for_device,
+	.map_sg			= swiotlb_map_sg_attrs,
+	.unmap_sg		= swiotlb_unmap_sg_attrs,
+	.sync_sg_for_cpu	= swiotlb_sync_sg_for_cpu,
+	.sync_sg_for_device	= swiotlb_sync_sg_for_device,
+	.dma_supported		= dmabounce_dma_supported,
+	.mapping_error		= dma_direct_mapping_error,
+};
+
+int brcm_pcie_bounce_init(struct device *dev,
+			  unsigned long buffer_size,
+			  dma_addr_t threshold)
+{
+	struct dmabounce_device_info *device_info;
+	int ret;
+
+	/* Only support a single client */
+	if (g_dmabounce_device_info)
+		return -EBUSY;
+
+	ret = bcm2838_dma40_memcpy_init();
+	if (ret)
+		return ret;
+
+	device_info = kmalloc(sizeof(struct dmabounce_device_info), GFP_ATOMIC);
+	if (!device_info) {
+		dev_err(dev,
+			"Could not allocated dmabounce_device_info\n");
+		return -ENOMEM;
+	}
+
+	ret = bounce_create(&device_info->pool, dev, buffer_size);
+	if (ret) {
+		dev_err(dev,
+			"dmabounce: could not allocate %ld byte DMA pool\n",
+			buffer_size);
+		goto err_bounce;
+	}
+
+	device_info->dev = dev;
+	device_info->threshold = threshold;
+	INIT_LIST_HEAD(&device_info->safe_buffers);
+	rwlock_init(&device_info->lock);
+
+	DO_STATS(device_info->map_count = 0);
+	DO_STATS(device_info->unmap_count = 0);
+	DO_STATS(device_info->sync_dev_count = 0);
+	DO_STATS(device_info->sync_cpu_count = 0);
+	DO_STATS(device_info->fail_count = 0);
+	DO_STATS(device_info->attr_res =
+		 device_create_file(dev, &dev_attr_dmabounce_stats));
+
+	g_dmabounce_device_info = device_info;
+
+	dev_err(dev, "dmabounce: initialised - %ld kB, threshold %pad\n",
+		 buffer_size / 1024, &threshold);
+
+	return 0;
+
+ err_bounce:
+	kfree(device_info);
+	return ret;
+}
+EXPORT_SYMBOL(brcm_pcie_bounce_init);
+
+void brcm_pcie_bounce_uninit(struct device *dev)
+{
+	struct dmabounce_device_info *device_info = g_dmabounce_device_info;
+
+	g_dmabounce_device_info = NULL;
+
+	if (!device_info) {
+		dev_warn(dev,
+			 "Never registered with dmabounce but attempting"
+			 "to unregister!\n");
+		return;
+	}
+
+	if (!list_empty(&device_info->safe_buffers)) {
+		dev_err(dev,
+			"Removing from dmabounce with pending buffers!\n");
+		BUG();
+	}
+
+	bounce_destroy(&device_info->pool);
+
+	DO_STATS(if (device_info->attr_res == 0)
+			 device_remove_file(dev, &dev_attr_dmabounce_stats));
+
+	kfree(device_info);
+}
+EXPORT_SYMBOL(brcm_pcie_bounce_uninit);
+
+int brcm_pcie_bounce_register_dev(struct device *dev)
+{
+	set_dma_ops(dev, &dmabounce_ops);
+
+	return 0;
+}
+EXPORT_SYMBOL(brcm_pcie_bounce_register_dev);
+
+MODULE_AUTHOR("Phil Elwell <phil@raspberrypi.org>");
+MODULE_DESCRIPTION("Dedicate DMA bounce support for pcie-brcmstb");
+MODULE_LICENSE("GPL");

drivers/pci/controller/pcie-brcmstb.c

@@ -617,28 +617,6 @@ static const struct dma_map_ops brcm_dma_ops = {
 
 static void brcm_set_dma_ops(struct device *dev)
 {
-	int ret;
-
-	if (IS_ENABLED(CONFIG_ARM64)) {
-		/*
-		 * We are going to invoke get_dma_ops().  That
-		 * function, at this point in time, invokes
-		 * get_arch_dma_ops(), and for ARM64 that function
-		 * returns a pointer to dummy_dma_ops.  So then we'd
-		 * like to call arch_setup_dma_ops(), but that isn't
-		 * exported.  Instead, we call of_dma_configure(),
-		 * which is exported, and this calls
-		 * arch_setup_dma_ops().  Once we do this the call to
-		 * get_dma_ops() will work properly because
-		 * dev->dma_ops will be set.
-		 */
-		ret = of_dma_configure(dev, dev->of_node, true);
-		if (ret) {
-			dev_err(dev, "of_dma_configure() failed: %d\n", ret);
-			return;
-		}
-	}
-
 	arch_dma_ops = get_dma_ops(dev);
 	if (!arch_dma_ops) {
 		dev_err(dev, "failed to get arch_dma_ops\n");
@@ -657,12 +635,12 @@ static int brcmstb_platform_notifier(struct notifier_block *nb,
 	extern unsigned long max_pfn;
 	struct device *dev = __dev;
 	const char *rc_name = "0000:00:00.0";
+	int ret;
 
 	switch (event) {
 	case BUS_NOTIFY_ADD_DEVICE:
 		if (max_pfn > (bounce_threshold/PAGE_SIZE) &&
 		    strcmp(dev->kobj.name, rc_name)) {
-			int ret;
 
 			ret = brcm_pcie_bounce_register_dev(dev);
 			if (ret) {
@@ -671,6 +649,12 @@ static int brcmstb_platform_notifier(struct notifier_block *nb,
 					ret);
 				return ret;
 			}
+		} else if (IS_ENABLED(CONFIG_ARM64)) {
+			ret = of_dma_configure(dev, dev->of_node, true);
+			if (ret) {
+				dev_err(dev, "of_dma_configure() failed: %d\n", ret);
+				return ret;
+			}
 		}
 		brcm_set_dma_ops(dev);
 		return NOTIFY_OK;