4 * Linux architectural port borrowing liberally from similar works of
5 * others. All original copyrights apply as per the original source
8 * Modifications for the OpenRISC architecture:
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.
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...
22 #include <linux/dma-mapping.h>
23 #include <linux/dma-debug.h>
24 #include <linux/export.h>
25 #include <linux/dma-attrs.h>
27 #include <asm/cpuinfo.h>
28 #include <asm/spr_defs.h>
29 #include <asm/tlbflush.h>
32 page_set_nocache(pte_t *pte, unsigned long addr,
33 unsigned long next, struct mm_walk *walk)
37 pte_val(*pte) |= _PAGE_CI;
40 * Flush the page out of the TLB so that the new page flags get
41 * picked up next time there's an access
43 flush_tlb_page(NULL, addr);
45 /* Flush page out of dcache */
46 for (cl = __pa(addr); cl < __pa(next); cl += cpuinfo.dcache_block_size)
53 page_clear_nocache(pte_t *pte, unsigned long addr,
54 unsigned long next, struct mm_walk *walk)
56 pte_val(*pte) &= ~_PAGE_CI;
59 * Flush the page out of the TLB so that the new page flags get
60 * picked up next time there's an access
62 flush_tlb_page(NULL, addr);
68 * Alloc "coherent" memory, which for OpenRISC means simply uncached.
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.
74 * If the NON_CONSISTENT attribute is set, then this function just
75 * returns "normal", cachable memory.
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.
84 or1k_dma_alloc(struct device *dev, size_t size,
85 dma_addr_t *dma_handle, gfp_t gfp,
86 struct dma_attrs *attrs)
90 struct mm_walk walk = {
91 .pte_entry = page_set_nocache,
95 page = alloc_pages_exact(size, gfp);
99 /* This gives us the real physical address of the first page. */
100 *dma_handle = __pa(page);
102 va = (unsigned long)page;
104 if (!dma_get_attr(DMA_ATTR_NON_CONSISTENT, attrs)) {
106 * We need to iterate through the pages, clearing the dcache for
107 * them and setting the cache-inhibit bit.
109 if (walk_page_range(va, va + size, &walk)) {
110 free_pages_exact(page, size);
119 or1k_dma_free(struct device *dev, size_t size, void *vaddr,
120 dma_addr_t dma_handle, struct dma_attrs *attrs)
122 unsigned long va = (unsigned long)vaddr;
123 struct mm_walk walk = {
124 .pte_entry = page_clear_nocache,
128 if (!dma_get_attr(DMA_ATTR_NON_CONSISTENT, attrs)) {
129 /* walk_page_range shouldn't be able to fail here */
130 WARN_ON(walk_page_range(va, va + size, &walk));
133 free_pages_exact(vaddr, size);
137 or1k_map_page(struct device *dev, struct page *page,
138 unsigned long offset, size_t size,
139 enum dma_data_direction dir,
140 struct dma_attrs *attrs)
143 dma_addr_t addr = page_to_phys(page) + offset;
147 /* Flush the dcache for the requested range */
148 for (cl = addr; cl < addr + size;
149 cl += cpuinfo.dcache_block_size)
150 mtspr(SPR_DCBFR, cl);
152 case DMA_FROM_DEVICE:
153 /* Invalidate the dcache for the requested range */
154 for (cl = addr; cl < addr + size;
155 cl += cpuinfo.dcache_block_size)
156 mtspr(SPR_DCBIR, cl);
160 * NOTE: If dir == DMA_BIDIRECTIONAL then there's no need to
161 * flush nor invalidate the cache here as the area will need
162 * to be manually synced anyway.
171 or1k_unmap_page(struct device *dev, dma_addr_t dma_handle,
172 size_t size, enum dma_data_direction dir,
173 struct dma_attrs *attrs)
175 /* Nothing special to do here... */
179 or1k_map_sg(struct device *dev, struct scatterlist *sg,
180 int nents, enum dma_data_direction dir,
181 struct dma_attrs *attrs)
183 struct scatterlist *s;
186 for_each_sg(sg, s, nents, i) {
187 s->dma_address = or1k_map_page(dev, sg_page(s), s->offset,
188 s->length, dir, NULL);
195 or1k_unmap_sg(struct device *dev, struct scatterlist *sg,
196 int nents, enum dma_data_direction dir,
197 struct dma_attrs *attrs)
199 struct scatterlist *s;
202 for_each_sg(sg, s, nents, i) {
203 or1k_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, NULL);
208 or1k_sync_single_for_cpu(struct device *dev,
209 dma_addr_t dma_handle, size_t size,
210 enum dma_data_direction dir)
213 dma_addr_t addr = dma_handle;
215 /* Invalidate the dcache for the requested range */
216 for (cl = addr; cl < addr + size; cl += cpuinfo.dcache_block_size)
217 mtspr(SPR_DCBIR, cl);
221 or1k_sync_single_for_device(struct device *dev,
222 dma_addr_t dma_handle, size_t size,
223 enum dma_data_direction dir)
226 dma_addr_t addr = dma_handle;
228 /* Flush the dcache for the requested range */
229 for (cl = addr; cl < addr + size; cl += cpuinfo.dcache_block_size)
230 mtspr(SPR_DCBFR, cl);
233 struct dma_map_ops or1k_dma_map_ops = {
234 .alloc = or1k_dma_alloc,
235 .free = or1k_dma_free,
236 .map_page = or1k_map_page,
237 .unmap_page = or1k_unmap_page,
238 .map_sg = or1k_map_sg,
239 .unmap_sg = or1k_unmap_sg,
240 .sync_single_for_cpu = or1k_sync_single_for_cpu,
241 .sync_single_for_device = or1k_sync_single_for_device,
243 EXPORT_SYMBOL(or1k_dma_map_ops);
245 /* Number of entries preallocated for DMA-API debugging */
246 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
248 static int __init dma_init(void)
250 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
254 fs_initcall(dma_init);