dma-noncoherent.c (9cdf083f981b8d37b3212400a359368661385099) | dma-noncoherent.c (f8c55dc6e828324fc58c0bb32d72a5a4041d1c3b) |
---|---|
1// SPDX-License-Identifier: GPL-2.0 |
|
1/* | 2/* |
2 * This file is subject to the terms and conditions of the GNU General Public 3 * License. See the file "COPYING" in the main directory of this archive 4 * for more details. 5 * | |
6 * Copyright (C) 2000 Ani Joshi <ajoshi@unixbox.com> | 3 * Copyright (C) 2000 Ani Joshi <ajoshi@unixbox.com> |
7 * Copyright (C) 2000, 2001 Ralf Baechle <ralf@gnu.org> | 4 * Copyright (C) 2000, 2001, 06 Ralf Baechle <ralf@linux-mips.org> |
8 * swiped from i386, and cloned for MIPS by Geert, polished by Ralf. 9 */ | 5 * swiped from i386, and cloned for MIPS by Geert, polished by Ralf. 6 */ |
10#include <linux/types.h> 11#include <linux/mm.h> 12#include <linux/module.h> 13#include <linux/string.h> 14#include <linux/dma-mapping.h> | 7#include <linux/dma-direct.h> 8#include <linux/dma-noncoherent.h> 9#include <linux/dma-contiguous.h> 10#include <linux/highmem.h> |
15 16#include <asm/cache.h> | 11 12#include <asm/cache.h> |
13#include <asm/cpu-type.h> 14#include <asm/dma-coherence.h> |
|
17#include <asm/io.h> 18 | 15#include <asm/io.h> 16 |
17#ifdef CONFIG_DMA_PERDEV_COHERENT 18static inline int dev_is_coherent(struct device *dev) 19{ 20 return dev->archdata.dma_coherent; 21} 22#else 23static inline int dev_is_coherent(struct device *dev) 24{ 25 switch (coherentio) { 26 default: 27 case IO_COHERENCE_DEFAULT: 28 return hw_coherentio; 29 case IO_COHERENCE_ENABLED: 30 return 1; 31 case IO_COHERENCE_DISABLED: 32 return 0; 33 } 34} 35#endif /* CONFIG_DMA_PERDEV_COHERENT */ 36 |
|
19/* | 37/* |
20 * Warning on the terminology - Linux calls an uncached area coherent; 21 * MIPS terminology calls memory areas with hardware maintained coherency 22 * coherent. | 38 * The affected CPUs below in 'cpu_needs_post_dma_flush()' can speculatively 39 * fill random cachelines with stale data at any time, requiring an extra 40 * flush post-DMA. 41 * 42 * Warning on the terminology - Linux calls an uncached area coherent; MIPS 43 * terminology calls memory areas with hardware maintained coherency coherent. 44 * 45 * Note that the R14000 and R16000 should also be checked for in this condition. 46 * However this function is only called on non-I/O-coherent systems and only the 47 * R10000 and R12000 are used in such systems, the SGI IP28 Indigo² rsp. 48 * SGI IP32 aka O2. |
23 */ | 49 */ |
24 25void *dma_alloc_noncoherent(struct device *dev, size_t size, 26 dma_addr_t * dma_handle, gfp_t gfp) | 50static inline bool cpu_needs_post_dma_flush(struct device *dev) |
27{ | 51{ |
28 void *ret; 29 /* ignore region specifiers */ 30 gfp &= ~(__GFP_DMA | __GFP_HIGHMEM); | 52 if (dev_is_coherent(dev)) 53 return false; |
31 | 54 |
32 if (dev == NULL || (dev->coherent_dma_mask < 0xffffffff)) 33 gfp |= GFP_DMA; 34 ret = (void *) __get_free_pages(gfp, get_order(size)); 35 36 if (ret != NULL) { 37 memset(ret, 0, size); 38 *dma_handle = virt_to_phys(ret); | 55 switch (boot_cpu_type()) { 56 case CPU_R10000: 57 case CPU_R12000: 58 case CPU_BMIPS5000: 59 return true; 60 default: 61 /* 62 * Presence of MAARs suggests that the CPU supports 63 * speculatively prefetching data, and therefore requires 64 * the post-DMA flush/invalidate. 65 */ 66 return cpu_has_maar; |
39 } | 67 } |
40 41 return ret; | |
42} 43 | 68} 69 |
44EXPORT_SYMBOL(dma_alloc_noncoherent); 45 46void *dma_alloc_coherent(struct device *dev, size_t size, 47 dma_addr_t * dma_handle, gfp_t gfp) | 70void *arch_dma_alloc(struct device *dev, size_t size, 71 dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs) |
48{ 49 void *ret; 50 | 72{ 73 void *ret; 74 |
51 ret = dma_alloc_noncoherent(dev, size, dma_handle, gfp); 52 if (ret) { | 75 ret = dma_direct_alloc(dev, size, dma_handle, gfp, attrs); 76 if (!ret) 77 return NULL; 78 79 if (!dev_is_coherent(dev) && !(attrs & DMA_ATTR_NON_CONSISTENT)) { |
53 dma_cache_wback_inv((unsigned long) ret, size); 54 ret = UNCAC_ADDR(ret); 55 } 56 57 return ret; 58} 59 | 80 dma_cache_wback_inv((unsigned long) ret, size); 81 ret = UNCAC_ADDR(ret); 82 } 83 84 return ret; 85} 86 |
60EXPORT_SYMBOL(dma_alloc_coherent); 61 62void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr, 63 dma_addr_t dma_handle) | 87void arch_dma_free(struct device *dev, size_t size, void *cpu_addr, 88 dma_addr_t dma_addr, unsigned long attrs) |
64{ | 89{ |
65 free_pages((unsigned long) vaddr, get_order(size)); | 90 if (!(attrs & DMA_ATTR_NON_CONSISTENT) && !dev_is_coherent(dev)) 91 cpu_addr = (void *)CAC_ADDR((unsigned long)cpu_addr); 92 dma_direct_free(dev, size, cpu_addr, dma_addr, attrs); |
66} 67 | 93} 94 |
68EXPORT_SYMBOL(dma_free_noncoherent); 69 70void dma_free_coherent(struct device *dev, size_t size, void *vaddr, 71 dma_addr_t dma_handle) | 95int arch_dma_mmap(struct device *dev, struct vm_area_struct *vma, 96 void *cpu_addr, dma_addr_t dma_addr, size_t size, 97 unsigned long attrs) |
72{ | 98{ |
73 unsigned long addr = (unsigned long) vaddr; | 99 unsigned long user_count = vma_pages(vma); 100 unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT; 101 unsigned long addr = (unsigned long)cpu_addr; 102 unsigned long off = vma->vm_pgoff; 103 unsigned long pfn; 104 int ret = -ENXIO; |
74 | 105 |
75 addr = CAC_ADDR(addr); 76 free_pages(addr, get_order(size)); 77} | 106 if (!dev_is_coherent(dev)) 107 addr = CAC_ADDR(addr); |
78 | 108 |
79EXPORT_SYMBOL(dma_free_coherent); | 109 pfn = page_to_pfn(virt_to_page((void *)addr)); |
80 | 110 |
81static inline void __dma_sync(unsigned long addr, size_t size, 82 enum dma_data_direction direction) | 111 if (attrs & DMA_ATTR_WRITE_COMBINE) 112 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot); 113 else 114 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 115 116 if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret)) 117 return ret; 118 119 if (off < count && user_count <= (count - off)) { 120 ret = remap_pfn_range(vma, vma->vm_start, 121 pfn + off, 122 user_count << PAGE_SHIFT, 123 vma->vm_page_prot); 124 } 125 126 return ret; 127} 128 129static inline void dma_sync_virt(void *addr, size_t size, 130 enum dma_data_direction dir) |
83{ | 131{ |
84 switch (direction) { | 132 switch (dir) { |
85 case DMA_TO_DEVICE: | 133 case DMA_TO_DEVICE: |
86 dma_cache_wback(addr, size); | 134 dma_cache_wback((unsigned long)addr, size); |
87 break; 88 89 case DMA_FROM_DEVICE: | 135 break; 136 137 case DMA_FROM_DEVICE: |
90 dma_cache_inv(addr, size); | 138 dma_cache_inv((unsigned long)addr, size); |
91 break; 92 93 case DMA_BIDIRECTIONAL: | 139 break; 140 141 case DMA_BIDIRECTIONAL: |
94 dma_cache_wback_inv(addr, size); | 142 dma_cache_wback_inv((unsigned long)addr, size); |
95 break; 96 97 default: 98 BUG(); 99 } 100} 101 | 143 break; 144 145 default: 146 BUG(); 147 } 148} 149 |
102dma_addr_t dma_map_single(struct device *dev, void *ptr, size_t size, 103 enum dma_data_direction direction) | 150/* 151 * A single sg entry may refer to multiple physically contiguous pages. But 152 * we still need to process highmem pages individually. If highmem is not 153 * configured then the bulk of this loop gets optimized out. 154 */ 155static inline void dma_sync_phys(phys_addr_t paddr, size_t size, 156 enum dma_data_direction dir) |
104{ | 157{ |
105 unsigned long addr = (unsigned long) ptr; | 158 struct page *page = pfn_to_page(paddr >> PAGE_SHIFT); 159 unsigned long offset = paddr & ~PAGE_MASK; 160 size_t left = size; |
106 | 161 |
107 __dma_sync(addr, size, direction); | 162 do { 163 size_t len = left; |
108 | 164 |
109 return virt_to_phys(ptr); 110} | 165 if (PageHighMem(page)) { 166 void *addr; |
111 | 167 |
112EXPORT_SYMBOL(dma_map_single); | 168 if (offset + len > PAGE_SIZE) { 169 if (offset >= PAGE_SIZE) { 170 page += offset >> PAGE_SHIFT; 171 offset &= ~PAGE_MASK; 172 } 173 len = PAGE_SIZE - offset; 174 } |
113 | 175 |
114void dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size, 115 enum dma_data_direction direction) 116{ 117 unsigned long addr; 118 addr = dma_addr + PAGE_OFFSET; 119 120 //__dma_sync(addr, size, direction); | 176 addr = kmap_atomic(page); 177 dma_sync_virt(addr + offset, len, dir); 178 kunmap_atomic(addr); 179 } else 180 dma_sync_virt(page_address(page) + offset, size, dir); 181 offset = 0; 182 page++; 183 left -= len; 184 } while (left); |
121} 122 | 185} 186 |
123EXPORT_SYMBOL(dma_unmap_single); 124 125int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents, 126 enum dma_data_direction direction) | 187void arch_sync_dma_for_device(struct device *dev, phys_addr_t paddr, 188 size_t size, enum dma_data_direction dir) |
127{ | 189{ |
128 int i; 129 130 BUG_ON(direction == DMA_NONE); 131 132 for (i = 0; i < nents; i++, sg++) { 133 unsigned long addr; 134 135 addr = (unsigned long) page_address(sg->page); 136 if (addr) { 137 __dma_sync(addr + sg->offset, sg->length, direction); 138 sg->dma_address = (dma_addr_t)page_to_phys(sg->page) 139 + sg->offset; 140 } 141 } 142 143 return nents; | 190 if (!dev_is_coherent(dev)) 191 dma_sync_phys(paddr, size, dir); |
144} 145 | 192} 193 |
146EXPORT_SYMBOL(dma_map_sg); 147 148dma_addr_t dma_map_page(struct device *dev, struct page *page, 149 unsigned long offset, size_t size, enum dma_data_direction direction) | 194void arch_sync_dma_for_cpu(struct device *dev, phys_addr_t paddr, 195 size_t size, enum dma_data_direction dir) |
150{ | 196{ |
151 unsigned long addr; 152 153 BUG_ON(direction == DMA_NONE); 154 155 addr = (unsigned long) page_address(page) + offset; 156 dma_cache_wback_inv(addr, size); 157 158 return page_to_phys(page) + offset; | 197 if (cpu_needs_post_dma_flush(dev)) 198 dma_sync_phys(paddr, size, dir); |
159} 160 | 199} 200 |
161EXPORT_SYMBOL(dma_map_page); 162 163void dma_unmap_page(struct device *dev, dma_addr_t dma_address, size_t size, 164 enum dma_data_direction direction) | 201void arch_dma_cache_sync(struct device *dev, void *vaddr, size_t size, 202 enum dma_data_direction direction) |
165{ 166 BUG_ON(direction == DMA_NONE); 167 | 203{ 204 BUG_ON(direction == DMA_NONE); 205 |
168 if (direction != DMA_TO_DEVICE) { 169 unsigned long addr; 170 171 addr = dma_address + PAGE_OFFSET; 172 dma_cache_wback_inv(addr, size); 173 } | 206 if (!dev_is_coherent(dev)) 207 dma_sync_virt(vaddr, size, direction); |
174} | 208} |
175 176EXPORT_SYMBOL(dma_unmap_page); 177 178void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nhwentries, 179 enum dma_data_direction direction) 180{ 181 unsigned long addr; 182 int i; 183 184 BUG_ON(direction == DMA_NONE); 185 186 if (direction == DMA_TO_DEVICE) 187 return; 188 189 for (i = 0; i < nhwentries; i++, sg++) { 190 addr = (unsigned long) page_address(sg->page); 191 if (addr) 192 __dma_sync(addr + sg->offset, sg->length, direction); 193 } 194} 195 196EXPORT_SYMBOL(dma_unmap_sg); 197 198void dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle, 199 size_t size, enum dma_data_direction direction) 200{ 201 unsigned long addr; 202 203 BUG_ON(direction == DMA_NONE); 204 205 addr = dma_handle + PAGE_OFFSET; 206 __dma_sync(addr, size, direction); 207} 208 209EXPORT_SYMBOL(dma_sync_single_for_cpu); 210 211void dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle, 212 size_t size, enum dma_data_direction direction) 213{ 214 unsigned long addr; 215 216 BUG_ON(direction == DMA_NONE); 217 218 addr = dma_handle + PAGE_OFFSET; 219 __dma_sync(addr, size, direction); 220} 221 222EXPORT_SYMBOL(dma_sync_single_for_device); 223 224void dma_sync_single_range_for_cpu(struct device *dev, dma_addr_t dma_handle, 225 unsigned long offset, size_t size, enum dma_data_direction direction) 226{ 227 unsigned long addr; 228 229 BUG_ON(direction == DMA_NONE); 230 231 addr = dma_handle + offset + PAGE_OFFSET; 232 __dma_sync(addr, size, direction); 233} 234 235EXPORT_SYMBOL(dma_sync_single_range_for_cpu); 236 237void dma_sync_single_range_for_device(struct device *dev, dma_addr_t dma_handle, 238 unsigned long offset, size_t size, enum dma_data_direction direction) 239{ 240 unsigned long addr; 241 242 BUG_ON(direction == DMA_NONE); 243 244 addr = dma_handle + offset + PAGE_OFFSET; 245 __dma_sync(addr, size, direction); 246} 247 248EXPORT_SYMBOL(dma_sync_single_range_for_device); 249 250void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg, int nelems, 251 enum dma_data_direction direction) 252{ 253 int i; 254 255 BUG_ON(direction == DMA_NONE); 256 257 /* Make sure that gcc doesn't leave the empty loop body. */ 258 for (i = 0; i < nelems; i++, sg++) 259 __dma_sync((unsigned long)page_address(sg->page), 260 sg->length, direction); 261} 262 263EXPORT_SYMBOL(dma_sync_sg_for_cpu); 264 265void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg, int nelems, 266 enum dma_data_direction direction) 267{ 268 int i; 269 270 BUG_ON(direction == DMA_NONE); 271 272 /* Make sure that gcc doesn't leave the empty loop body. */ 273 for (i = 0; i < nelems; i++, sg++) 274 __dma_sync((unsigned long)page_address(sg->page), 275 sg->length, direction); 276} 277 278EXPORT_SYMBOL(dma_sync_sg_for_device); 279 280int dma_mapping_error(dma_addr_t dma_addr) 281{ 282 return 0; 283} 284 285EXPORT_SYMBOL(dma_mapping_error); 286 287int dma_supported(struct device *dev, u64 mask) 288{ 289 /* 290 * we fall back to GFP_DMA when the mask isn't all 1s, 291 * so we can't guarantee allocations that must be 292 * within a tighter range than GFP_DMA.. 293 */ 294 if (mask < 0x00ffffff) 295 return 0; 296 297 return 1; 298} 299 300EXPORT_SYMBOL(dma_supported); 301 302int dma_is_consistent(struct device *dev, dma_addr_t dma_addr) 303{ 304 return 1; 305} 306 307EXPORT_SYMBOL(dma_is_consistent); 308 309void dma_cache_sync(struct device *dev, void *vaddr, size_t size, 310 enum dma_data_direction direction) 311{ 312 if (direction == DMA_NONE) 313 return; 314 315 dma_cache_wback_inv((unsigned long)vaddr, size); 316} 317 318EXPORT_SYMBOL(dma_cache_sync); 319 320/* The DAC routines are a PCIism.. */ 321 322#ifdef CONFIG_PCI 323 324#include <linux/pci.h> 325 326dma64_addr_t pci_dac_page_to_dma(struct pci_dev *pdev, 327 struct page *page, unsigned long offset, int direction) 328{ 329 return (dma64_addr_t)page_to_phys(page) + offset; 330} 331 332EXPORT_SYMBOL(pci_dac_page_to_dma); 333 334struct page *pci_dac_dma_to_page(struct pci_dev *pdev, 335 dma64_addr_t dma_addr) 336{ 337 return mem_map + (dma_addr >> PAGE_SHIFT); 338} 339 340EXPORT_SYMBOL(pci_dac_dma_to_page); 341 342unsigned long pci_dac_dma_to_offset(struct pci_dev *pdev, 343 dma64_addr_t dma_addr) 344{ 345 return dma_addr & ~PAGE_MASK; 346} 347 348EXPORT_SYMBOL(pci_dac_dma_to_offset); 349 350void pci_dac_dma_sync_single_for_cpu(struct pci_dev *pdev, 351 dma64_addr_t dma_addr, size_t len, int direction) 352{ 353 BUG_ON(direction == PCI_DMA_NONE); 354 355 dma_cache_wback_inv(dma_addr + PAGE_OFFSET, len); 356} 357 358EXPORT_SYMBOL(pci_dac_dma_sync_single_for_cpu); 359 360void pci_dac_dma_sync_single_for_device(struct pci_dev *pdev, 361 dma64_addr_t dma_addr, size_t len, int direction) 362{ 363 BUG_ON(direction == PCI_DMA_NONE); 364 365 dma_cache_wback_inv(dma_addr + PAGE_OFFSET, len); 366} 367 368EXPORT_SYMBOL(pci_dac_dma_sync_single_for_device); 369 370#endif /* CONFIG_PCI */ | |