1/* 2 * Copyright (C) 2006 Benjamin Herrenschmidt, IBM Corporation 3 * 4 * Provide default implementations of the DMA mapping callbacks for 5 * directly mapped busses. 6 */ 7 8#include <linux/device.h> 9#include <linux/dma-mapping.h> 10#include <linux/dma-debug.h> 11#include <linux/gfp.h> 12#include <linux/memblock.h> 13#include <linux/export.h> 14#include <linux/pci.h> 15#include <asm/vio.h> 16#include <asm/bug.h> 17#include <asm/machdep.h> 18#include <asm/swiotlb.h> 19 20/* 21 * Generic direct DMA implementation 22 * 23 * This implementation supports a per-device offset that can be applied if 24 * the address at which memory is visible to devices is not 0. Platform code 25 * can set archdata.dma_data to an unsigned long holding the offset. By 26 * default the offset is PCI_DRAM_OFFSET. 27 */ 28 29static u64 __maybe_unused get_pfn_limit(struct device *dev) 30{ 31 u64 pfn = (dev->coherent_dma_mask >> PAGE_SHIFT) + 1; 32 struct dev_archdata __maybe_unused *sd = &dev->archdata; 33 34#ifdef CONFIG_SWIOTLB 35 if (sd->max_direct_dma_addr && sd->dma_ops == &swiotlb_dma_ops) 36 pfn = min_t(u64, pfn, sd->max_direct_dma_addr >> PAGE_SHIFT); 37#endif 38 39 return pfn; 40} 41 42void *dma_direct_alloc_coherent(struct device *dev, size_t size, 43 dma_addr_t *dma_handle, gfp_t flag, 44 struct dma_attrs *attrs) 45{ 46 void *ret; 47#ifdef CONFIG_NOT_COHERENT_CACHE 48 ret = __dma_alloc_coherent(dev, size, dma_handle, flag); 49 if (ret == NULL) 50 return NULL; 51 *dma_handle += get_dma_offset(dev); 52 return ret; 53#else 54 struct page *page; 55 int node = dev_to_node(dev); 56#ifdef CONFIG_FSL_SOC 57 u64 pfn = get_pfn_limit(dev); 58 int zone; 59 60 /* 61 * This code should be OK on other platforms, but we have drivers that 62 * don't set coherent_dma_mask. As a workaround we just ifdef it. This 63 * whole routine needs some serious cleanup. 64 */ 65 66 zone = dma_pfn_limit_to_zone(pfn); 67 if (zone < 0) { 68 dev_err(dev, "%s: No suitable zone for pfn %#llx\n", 69 __func__, pfn); 70 return NULL; 71 } 72 73 switch (zone) { 74 case ZONE_DMA: 75 flag |= GFP_DMA; 76 break; 77#ifdef CONFIG_ZONE_DMA32 78 case ZONE_DMA32: 79 flag |= GFP_DMA32; 80 break; 81#endif 82 }; 83#endif /* CONFIG_FSL_SOC */ 84 85 /* ignore region specifiers */ 86 flag &= ~(__GFP_HIGHMEM); 87 88 page = alloc_pages_node(node, flag, get_order(size)); 89 if (page == NULL) 90 return NULL; 91 ret = page_address(page); 92 memset(ret, 0, size); 93 *dma_handle = __pa(ret) + get_dma_offset(dev); 94 95 return ret; 96#endif 97} 98 99void dma_direct_free_coherent(struct device *dev, size_t size, 100 void *vaddr, dma_addr_t dma_handle, 101 struct dma_attrs *attrs) 102{ 103#ifdef CONFIG_NOT_COHERENT_CACHE 104 __dma_free_coherent(size, vaddr); 105#else 106 free_pages((unsigned long)vaddr, get_order(size)); 107#endif 108} 109 110int dma_direct_mmap_coherent(struct device *dev, struct vm_area_struct *vma, 111 void *cpu_addr, dma_addr_t handle, size_t size, 112 struct dma_attrs *attrs) 113{ 114 unsigned long pfn; 115 116#ifdef CONFIG_NOT_COHERENT_CACHE 117 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 118 pfn = __dma_get_coherent_pfn((unsigned long)cpu_addr); 119#else 120 pfn = page_to_pfn(virt_to_page(cpu_addr)); 121#endif 122 return remap_pfn_range(vma, vma->vm_start, 123 pfn + vma->vm_pgoff, 124 vma->vm_end - vma->vm_start, 125 vma->vm_page_prot); 126} 127 128static int dma_direct_map_sg(struct device *dev, struct scatterlist *sgl, 129 int nents, enum dma_data_direction direction, 130 struct dma_attrs *attrs) 131{ 132 struct scatterlist *sg; 133 int i; 134 135 for_each_sg(sgl, sg, nents, i) { 136 sg->dma_address = sg_phys(sg) + get_dma_offset(dev); 137 sg->dma_length = sg->length; 138 __dma_sync_page(sg_page(sg), sg->offset, sg->length, direction); 139 } 140 141 return nents; 142} 143 144static void dma_direct_unmap_sg(struct device *dev, struct scatterlist *sg, 145 int nents, enum dma_data_direction direction, 146 struct dma_attrs *attrs) 147{ 148} 149 150static int dma_direct_dma_supported(struct device *dev, u64 mask) 151{ 152#ifdef CONFIG_PPC64 153 /* Could be improved so platforms can set the limit in case 154 * they have limited DMA windows 155 */ 156 return mask >= get_dma_offset(dev) + (memblock_end_of_DRAM() - 1); 157#else 158 return 1; 159#endif 160} 161 162static u64 dma_direct_get_required_mask(struct device *dev) 163{ 164 u64 end, mask; 165 166 end = memblock_end_of_DRAM() + get_dma_offset(dev); 167 168 mask = 1ULL << (fls64(end) - 1); 169 mask += mask - 1; 170 171 return mask; 172} 173 174static inline dma_addr_t dma_direct_map_page(struct device *dev, 175 struct page *page, 176 unsigned long offset, 177 size_t size, 178 enum dma_data_direction dir, 179 struct dma_attrs *attrs) 180{ 181 BUG_ON(dir == DMA_NONE); 182 __dma_sync_page(page, offset, size, dir); 183 return page_to_phys(page) + offset + get_dma_offset(dev); 184} 185 186static inline void dma_direct_unmap_page(struct device *dev, 187 dma_addr_t dma_address, 188 size_t size, 189 enum dma_data_direction direction, 190 struct dma_attrs *attrs) 191{ 192} 193 194#ifdef CONFIG_NOT_COHERENT_CACHE 195static inline void dma_direct_sync_sg(struct device *dev, 196 struct scatterlist *sgl, int nents, 197 enum dma_data_direction direction) 198{ 199 struct scatterlist *sg; 200 int i; 201 202 for_each_sg(sgl, sg, nents, i) 203 __dma_sync_page(sg_page(sg), sg->offset, sg->length, direction); 204} 205 206static inline void dma_direct_sync_single(struct device *dev, 207 dma_addr_t dma_handle, size_t size, 208 enum dma_data_direction direction) 209{ 210 __dma_sync(bus_to_virt(dma_handle), size, direction); 211} 212#endif 213 214struct dma_map_ops dma_direct_ops = { 215 .alloc = dma_direct_alloc_coherent, 216 .free = dma_direct_free_coherent, 217 .mmap = dma_direct_mmap_coherent, 218 .map_sg = dma_direct_map_sg, 219 .unmap_sg = dma_direct_unmap_sg, 220 .dma_supported = dma_direct_dma_supported, 221 .map_page = dma_direct_map_page, 222 .unmap_page = dma_direct_unmap_page, 223 .get_required_mask = dma_direct_get_required_mask, 224#ifdef CONFIG_NOT_COHERENT_CACHE 225 .sync_single_for_cpu = dma_direct_sync_single, 226 .sync_single_for_device = dma_direct_sync_single, 227 .sync_sg_for_cpu = dma_direct_sync_sg, 228 .sync_sg_for_device = dma_direct_sync_sg, 229#endif 230}; 231EXPORT_SYMBOL(dma_direct_ops); 232 233#define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16) 234 235int __dma_set_mask(struct device *dev, u64 dma_mask) 236{ 237 struct dma_map_ops *dma_ops = get_dma_ops(dev); 238 239 if ((dma_ops != NULL) && (dma_ops->set_dma_mask != NULL)) 240 return dma_ops->set_dma_mask(dev, dma_mask); 241 if (!dev->dma_mask || !dma_supported(dev, dma_mask)) 242 return -EIO; 243 *dev->dma_mask = dma_mask; 244 return 0; 245} 246 247int dma_set_mask(struct device *dev, u64 dma_mask) 248{ 249 if (ppc_md.dma_set_mask) 250 return ppc_md.dma_set_mask(dev, dma_mask); 251 return __dma_set_mask(dev, dma_mask); 252} 253EXPORT_SYMBOL(dma_set_mask); 254 255u64 __dma_get_required_mask(struct device *dev) 256{ 257 struct dma_map_ops *dma_ops = get_dma_ops(dev); 258 259 if (unlikely(dma_ops == NULL)) 260 return 0; 261 262 if (dma_ops->get_required_mask) 263 return dma_ops->get_required_mask(dev); 264 265 return DMA_BIT_MASK(8 * sizeof(dma_addr_t)); 266} 267 268u64 dma_get_required_mask(struct device *dev) 269{ 270 if (ppc_md.dma_get_required_mask) 271 return ppc_md.dma_get_required_mask(dev); 272 273 return __dma_get_required_mask(dev); 274} 275EXPORT_SYMBOL_GPL(dma_get_required_mask); 276 277static int __init dma_init(void) 278{ 279 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES); 280#ifdef CONFIG_PCI 281 dma_debug_add_bus(&pci_bus_type); 282#endif 283#ifdef CONFIG_IBMVIO 284 dma_debug_add_bus(&vio_bus_type); 285#endif 286 287 return 0; 288} 289fs_initcall(dma_init); 290 291