Hi, David,
Seems like you are the original author of code in
arch/mips/cavium-octeon/dma-octeon.c. Could you please tell me why we
need mb() in alloc_coherent(), map_page(), map_sg()? It seems like
because of cache coherency (CPU write some data, then map the page for
a device, if without mb(), then device may read wrong data.) but I'm
not sure.
On Tue, Aug 14, 2012 at 1:54 AM, Konrad Rzeszutek Wilk
<konrad.wilk@xxxxxxxxxx> wrote:
>> +static void *loongson_dma_alloc_coherent(struct device *dev, size_t size,
>> + dma_addr_t *dma_handle, gfp_t gfp, struct dma_attrs *attrs)
>> +{
>> + void *ret;
>> +
>> + if (dma_alloc_from_coherent(dev, size, dma_handle, &ret))
>> + return ret;
>> +
>> + /* ignore region specifiers */
>> + gfp &= ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM);
>> +
>> +#ifdef CONFIG_ZONE_DMA
>> + if (dev == NULL)
>> + gfp |= __GFP_DMA;
>
> When would this happen? dev == NULL?
>
>> + else if (dev->coherent_dma_mask <= DMA_BIT_MASK(24))
>> + gfp |= __GFP_DMA;
>> + else
>> +#endif
>> +#ifdef CONFIG_ZONE_DMA32
>> + if (dev->coherent_dma_mask <= DMA_BIT_MASK(32))
>> + gfp |= __GFP_DMA32;
>> + else
>
> Why the 'else'
>> +#endif
>> + ;
>
> why?
>> + gfp |= __GFP_NORETRY;
>> +
>> + ret = swiotlb_alloc_coherent(dev, size, dma_handle, gfp);
>> + mb();
>
> Why the 'mb()' ? Can you just do
> return swiotlb_alloc_coherent(...)
>
>> + return ret;
>> +}
>> +
>> +static void loongson_dma_free_coherent(struct device *dev, size_t size,
>> + void *vaddr, dma_addr_t dma_handle, struct dma_attrs *attrs)
>> +{
>> + int order = get_order(size);
>> +
>> + if (dma_release_from_coherent(dev, order, vaddr))
>> + return;
>> +
>> + swiotlb_free_coherent(dev, size, vaddr, dma_handle);
>> +}
>> +
>> +static dma_addr_t loongson_dma_map_page(struct device *dev, struct page *page,
>> + unsigned long offset, size_t size,
>> + enum dma_data_direction dir,
>> + struct dma_attrs *attrs)
>> +{
>> + dma_addr_t daddr = swiotlb_map_page(dev, page, offset, size,
>> + dir, attrs);
>> + mb();
>
> Please do 'return swiotlb_map_page(..)'..
>
> But if you are doing that why don't you just set the dma_ops.map_page = swiotlb_map_page
> ?
>
>
>> + return daddr;
>> +}
>> +
>> +static int loongson_dma_map_sg(struct device *dev, struct scatterlist *sg,
>> + int nents, enum dma_data_direction dir,
>> + struct dma_attrs *attrs)
>> +{
>> + int r = swiotlb_map_sg_attrs(dev, sg, nents, dir, NULL);
>> + mb();
>> +
>> + return r;
>> +}
>> +
>> +static void loongson_dma_sync_single_for_device(struct device *dev,
>> + dma_addr_t dma_handle, size_t size,
>> + enum dma_data_direction dir)
>> +{
>> + swiotlb_sync_single_for_device(dev, dma_handle, size, dir);
>> + mb();
>> +}
>> +
>> +static void loongson_dma_sync_sg_for_device(struct device *dev,
>> + struct scatterlist *sg, int nents,
>> + enum dma_data_direction dir)
>> +{
>> + swiotlb_sync_sg_for_device(dev, sg, nents, dir);
>> + mb();
>> +}
>> +
>
> I am not really sure why you have these extra functions, when you could
> just modify the dma_ops to point to the swiotlb ones
>
>> +static dma_addr_t loongson_unity_phys_to_dma(struct device *dev, phys_addr_t paddr)
>> +{
>> + return (paddr < 0x10000000) ?
>> + (paddr | 0x0000000080000000) : paddr;
>> +}
>> +
>> +static phys_addr_t loongson_unity_dma_to_phys(struct device *dev, dma_addr_t daddr)
>> +{
>> + return (daddr < 0x90000000 && daddr >= 0x80000000) ?
>> + (daddr & 0x0fffffff) : daddr;
>> +}
>> +
>> +struct loongson_dma_map_ops {
>> + struct dma_map_ops dma_map_ops;
>> + dma_addr_t (*phys_to_dma)(struct device *dev, phys_addr_t paddr);
>> + phys_addr_t (*dma_to_phys)(struct device *dev, dma_addr_t daddr);
>> +};
>> +
>> +dma_addr_t phys_to_dma(struct device *dev, phys_addr_t paddr)
>> +{
>> + struct loongson_dma_map_ops *ops = container_of(get_dma_ops(dev),
>> + struct loongson_dma_map_ops, dma_map_ops);
>> +
>> + return ops->phys_to_dma(dev, paddr);
>> +}
>> +
>> +phys_addr_t dma_to_phys(struct device *dev, dma_addr_t daddr)
>> +{
>> + struct loongson_dma_map_ops *ops = container_of(get_dma_ops(dev),
>> + struct loongson_dma_map_ops, dma_map_ops);
>> +
>> + return ops->dma_to_phys(dev, daddr);
>> +}
>> +
>> +static int loongson_dma_set_mask(struct device *dev, u64 mask)
>> +{
>> + /* Loongson doesn't support DMA above 32-bit */
>> + if (mask > DMA_BIT_MASK(32))
>> + return -EIO;
>> +
>> + *dev->dma_mask = mask;
>> +
>> + return 0;
>> +}
>> +
>> +static struct loongson_dma_map_ops loongson_linear_dma_map_ops = {
>> + .dma_map_ops = {
>> + .alloc = loongson_dma_alloc_coherent,
>> + .free = loongson_dma_free_coherent,
>> + .map_page = loongson_dma_map_page,
>
> But why not 'swiotlb_map_page'?
>
>> + .unmap_page = swiotlb_unmap_page,
>> + .map_sg = loongson_dma_map_sg,
>> + .unmap_sg = swiotlb_unmap_sg_attrs,
>> + .sync_single_for_cpu = swiotlb_sync_single_for_cpu,
>> + .sync_single_for_device = loongson_dma_sync_single_for_device,
>> + .sync_sg_for_cpu = swiotlb_sync_sg_for_cpu,
>> + .sync_sg_for_device = loongson_dma_sync_sg_for_device,
>> + .mapping_error = swiotlb_dma_mapping_error,
>> + .dma_supported = swiotlb_dma_supported,
>> + .set_dma_mask = loongson_dma_set_mask
>> + },
>> + .phys_to_dma = loongson_unity_phys_to_dma,
>> + .dma_to_phys = loongson_unity_dma_to_phys
>
> Why do you need these? I am not seeing it being used here by any external code?
>
>> +};
>> +
>> +void __init plat_swiotlb_setup(void)
>> +{
>> + swiotlb_init(1);
>> + mips_dma_map_ops = &loongson_linear_dma_map_ops.dma_map_ops;
>> +}
>> diff --git a/arch/mips/mm/dma-default.c b/arch/mips/mm/dma-default.c
>> index 3fab204..122f4f8 100644
>> --- a/arch/mips/mm/dma-default.c
>> +++ b/arch/mips/mm/dma-default.c
>> @@ -42,6 +42,13 @@ static inline int cpu_is_noncoherent_r10000(struct device *dev)
>> current_cpu_type() == CPU_R12000);
>> }
>>
>> +static inline int cpu_is_noncoherent_loongson(struct device *dev)
>> +{
>> + return !plat_device_is_coherent(dev) &&
>> + (current_cpu_type() == CPU_LOONGSON2 ||
>> + current_cpu_type() == CPU_LOONGSON3);
>> +}
>> +
>> static gfp_t massage_gfp_flags(const struct device *dev, gfp_t gfp)
>> {
>> gfp_t dma_flag;
>> @@ -209,7 +216,7 @@ static inline void __dma_sync(struct page *page,
>> static void mips_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
>> size_t size, enum dma_data_direction direction, struct dma_attrs *attrs)
>> {
>> - if (cpu_is_noncoherent_r10000(dev))
>> + if (cpu_is_noncoherent_r10000(dev) || cpu_is_noncoherent_loongson(dev))
>> __dma_sync(dma_addr_to_page(dev, dma_addr),
>> dma_addr & ~PAGE_MASK, size, direction);
>>
>> @@ -260,7 +267,7 @@ static void mips_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
>> static void mips_dma_sync_single_for_cpu(struct device *dev,
>> dma_addr_t dma_handle, size_t size, enum dma_data_direction direction)
>> {
>> - if (cpu_is_noncoherent_r10000(dev))
>> + if (cpu_is_noncoherent_r10000(dev) || cpu_is_noncoherent_loongson(dev))
>> __dma_sync(dma_addr_to_page(dev, dma_handle),
>> dma_handle & ~PAGE_MASK, size, direction);
>> }
>> @@ -281,7 +288,7 @@ static void mips_dma_sync_sg_for_cpu(struct device *dev,
>>
>> /* Make sure that gcc doesn't leave the empty loop body. */
>> for (i = 0; i < nelems; i++, sg++) {
>> - if (cpu_is_noncoherent_r10000(dev))
>> + if (cpu_is_noncoherent_r10000(dev) || cpu_is_noncoherent_loongson(dev))
>> __dma_sync(sg_page(sg), sg->offset, sg->length,
>> direction);
>> }
>> --
>> 1.7.7.3
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>> the body of a message to majordomo@xxxxxxxxxxxxxxx
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>> Please read the FAQ at http://www.tux.org/lkml/
