1 /* 2 * OpenRISC Linux 3 * 4 * Linux architectural port borrowing liberally from similar works of 5 * others. All original copyrights apply as per the original source 6 * declaration. 7 * 8 * Modifications for the OpenRISC architecture: 9 * Copyright (C) 2003 Matjaz Breskvar <phoenix@bsemi.com> 10 * Copyright (C) 2010-2011 Jonas Bonn <jonas@southpole.se> 11 * 12 * This program is free software; you can redistribute it and/or 13 * modify it under the terms of the GNU General Public License 14 * as published by the Free Software Foundation; either version 15 * 2 of the License, or (at your option) any later version. 16 * 17 * DMA mapping callbacks... 18 * As alloc_coherent is the only DMA callback being used currently, that's 19 * the only thing implemented properly. The rest need looking into... 20 */ 21 22 #include <linux/dma-mapping.h> 23 #include <linux/dma-debug.h> 24 #include <linux/export.h> 25 26 #include <asm/cpuinfo.h> 27 #include <asm/spr_defs.h> 28 #include <asm/tlbflush.h> 29 30 static int 31 page_set_nocache(pte_t *pte, unsigned long addr, 32 unsigned long next, struct mm_walk *walk) 33 { 34 unsigned long cl; 35 struct cpuinfo_or1k *cpuinfo = &cpuinfo_or1k[smp_processor_id()]; 36 37 pte_val(*pte) |= _PAGE_CI; 38 39 /* 40 * Flush the page out of the TLB so that the new page flags get 41 * picked up next time there's an access 42 */ 43 flush_tlb_page(NULL, addr); 44 45 /* Flush page out of dcache */ 46 for (cl = __pa(addr); cl < __pa(next); cl += cpuinfo->dcache_block_size) 47 mtspr(SPR_DCBFR, cl); 48 49 return 0; 50 } 51 52 static int 53 page_clear_nocache(pte_t *pte, unsigned long addr, 54 unsigned long next, struct mm_walk *walk) 55 { 56 pte_val(*pte) &= ~_PAGE_CI; 57 58 /* 59 * Flush the page out of the TLB so that the new page flags get 60 * picked up next time there's an access 61 */ 62 flush_tlb_page(NULL, addr); 63 64 return 0; 65 } 66 67 /* 68 * Alloc "coherent" memory, which for OpenRISC means simply uncached. 69 * 70 * This function effectively just calls __get_free_pages, sets the 71 * cache-inhibit bit on those pages, and makes sure that the pages are 72 * flushed out of the cache before they are used. 73 * 74 * If the NON_CONSISTENT attribute is set, then this function just 75 * returns "normal", cachable memory. 76 * 77 * There are additional flags WEAK_ORDERING and WRITE_COMBINE to take 78 * into consideration here, too. All current known implementations of 79 * the OR1K support only strongly ordered memory accesses, so that flag 80 * is being ignored for now; uncached but write-combined memory is a 81 * missing feature of the OR1K. 82 */ 83 static void * 84 or1k_dma_alloc(struct device *dev, size_t size, 85 dma_addr_t *dma_handle, gfp_t gfp, 86 unsigned long attrs) 87 { 88 unsigned long va; 89 void *page; 90 struct mm_walk walk = { 91 .pte_entry = page_set_nocache, 92 .mm = &init_mm 93 }; 94 95 page = alloc_pages_exact(size, gfp); 96 if (!page) 97 return NULL; 98 99 /* This gives us the real physical address of the first page. */ 100 *dma_handle = __pa(page); 101 102 va = (unsigned long)page; 103 104 if ((attrs & DMA_ATTR_NON_CONSISTENT) == 0) { 105 /* 106 * We need to iterate through the pages, clearing the dcache for 107 * them and setting the cache-inhibit bit. 108 */ 109 if (walk_page_range(va, va + size, &walk)) { 110 free_pages_exact(page, size); 111 return NULL; 112 } 113 } 114 115 return (void *)va; 116 } 117 118 static void 119 or1k_dma_free(struct device *dev, size_t size, void *vaddr, 120 dma_addr_t dma_handle, unsigned long attrs) 121 { 122 unsigned long va = (unsigned long)vaddr; 123 struct mm_walk walk = { 124 .pte_entry = page_clear_nocache, 125 .mm = &init_mm 126 }; 127 128 if ((attrs & DMA_ATTR_NON_CONSISTENT) == 0) { 129 /* walk_page_range shouldn't be able to fail here */ 130 WARN_ON(walk_page_range(va, va + size, &walk)); 131 } 132 133 free_pages_exact(vaddr, size); 134 } 135 136 static dma_addr_t 137 or1k_map_page(struct device *dev, struct page *page, 138 unsigned long offset, size_t size, 139 enum dma_data_direction dir, 140 unsigned long attrs) 141 { 142 unsigned long cl; 143 dma_addr_t addr = page_to_phys(page) + offset; 144 struct cpuinfo_or1k *cpuinfo = &cpuinfo_or1k[smp_processor_id()]; 145 146 if (attrs & DMA_ATTR_SKIP_CPU_SYNC) 147 return addr; 148 149 switch (dir) { 150 case DMA_TO_DEVICE: 151 /* Flush the dcache for the requested range */ 152 for (cl = addr; cl < addr + size; 153 cl += cpuinfo->dcache_block_size) 154 mtspr(SPR_DCBFR, cl); 155 break; 156 case DMA_FROM_DEVICE: 157 /* Invalidate the dcache for the requested range */ 158 for (cl = addr; cl < addr + size; 159 cl += cpuinfo->dcache_block_size) 160 mtspr(SPR_DCBIR, cl); 161 break; 162 default: 163 /* 164 * NOTE: If dir == DMA_BIDIRECTIONAL then there's no need to 165 * flush nor invalidate the cache here as the area will need 166 * to be manually synced anyway. 167 */ 168 break; 169 } 170 171 return addr; 172 } 173 174 static void 175 or1k_unmap_page(struct device *dev, dma_addr_t dma_handle, 176 size_t size, enum dma_data_direction dir, 177 unsigned long attrs) 178 { 179 /* Nothing special to do here... */ 180 } 181 182 static int 183 or1k_map_sg(struct device *dev, struct scatterlist *sg, 184 int nents, enum dma_data_direction dir, 185 unsigned long attrs) 186 { 187 struct scatterlist *s; 188 int i; 189 190 for_each_sg(sg, s, nents, i) { 191 s->dma_address = or1k_map_page(dev, sg_page(s), s->offset, 192 s->length, dir, 0); 193 } 194 195 return nents; 196 } 197 198 static void 199 or1k_unmap_sg(struct device *dev, struct scatterlist *sg, 200 int nents, enum dma_data_direction dir, 201 unsigned long attrs) 202 { 203 struct scatterlist *s; 204 int i; 205 206 for_each_sg(sg, s, nents, i) { 207 or1k_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, 0); 208 } 209 } 210 211 static void 212 or1k_sync_single_for_cpu(struct device *dev, 213 dma_addr_t dma_handle, size_t size, 214 enum dma_data_direction dir) 215 { 216 unsigned long cl; 217 dma_addr_t addr = dma_handle; 218 struct cpuinfo_or1k *cpuinfo = &cpuinfo_or1k[smp_processor_id()]; 219 220 /* Invalidate the dcache for the requested range */ 221 for (cl = addr; cl < addr + size; cl += cpuinfo->dcache_block_size) 222 mtspr(SPR_DCBIR, cl); 223 } 224 225 static void 226 or1k_sync_single_for_device(struct device *dev, 227 dma_addr_t dma_handle, size_t size, 228 enum dma_data_direction dir) 229 { 230 unsigned long cl; 231 dma_addr_t addr = dma_handle; 232 struct cpuinfo_or1k *cpuinfo = &cpuinfo_or1k[smp_processor_id()]; 233 234 /* Flush the dcache for the requested range */ 235 for (cl = addr; cl < addr + size; cl += cpuinfo->dcache_block_size) 236 mtspr(SPR_DCBFR, cl); 237 } 238 239 const struct dma_map_ops or1k_dma_map_ops = { 240 .alloc = or1k_dma_alloc, 241 .free = or1k_dma_free, 242 .map_page = or1k_map_page, 243 .unmap_page = or1k_unmap_page, 244 .map_sg = or1k_map_sg, 245 .unmap_sg = or1k_unmap_sg, 246 .sync_single_for_cpu = or1k_sync_single_for_cpu, 247 .sync_single_for_device = or1k_sync_single_for_device, 248 }; 249 EXPORT_SYMBOL(or1k_dma_map_ops); 250 251 /* Number of entries preallocated for DMA-API debugging */ 252 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16) 253 254 static int __init dma_init(void) 255 { 256 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES); 257 258 return 0; 259 } 260 fs_initcall(dma_init); 261