Michal Simek | 3a0d7a4 | 2010-02-22 12:16:08 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Microblaze support for cache consistent memory. |
| 3 | * Copyright (C) 2010 Michal Simek <monstr@monstr.eu> |
| 4 | * Copyright (C) 2010 PetaLogix |
| 5 | * Copyright (C) 2005 John Williams <jwilliams@itee.uq.edu.au> |
| 6 | * |
| 7 | * Based on PowerPC version derived from arch/arm/mm/consistent.c |
| 8 | * Copyright (C) 2001 Dan Malek (dmalek@jlc.net) |
| 9 | * Copyright (C) 2000 Russell King |
| 10 | * |
| 11 | * This program is free software; you can redistribute it and/or modify |
| 12 | * it under the terms of the GNU General Public License version 2 as |
| 13 | * published by the Free Software Foundation. |
| 14 | */ |
| 15 | |
| 16 | #include <linux/module.h> |
| 17 | #include <linux/signal.h> |
| 18 | #include <linux/sched.h> |
| 19 | #include <linux/kernel.h> |
| 20 | #include <linux/errno.h> |
| 21 | #include <linux/string.h> |
| 22 | #include <linux/types.h> |
| 23 | #include <linux/ptrace.h> |
| 24 | #include <linux/mman.h> |
| 25 | #include <linux/mm.h> |
| 26 | #include <linux/swap.h> |
| 27 | #include <linux/stddef.h> |
| 28 | #include <linux/vmalloc.h> |
| 29 | #include <linux/init.h> |
| 30 | #include <linux/delay.h> |
| 31 | #include <linux/bootmem.h> |
| 32 | #include <linux/highmem.h> |
| 33 | #include <linux/pci.h> |
| 34 | #include <linux/interrupt.h> |
| 35 | |
| 36 | #include <asm/pgalloc.h> |
| 37 | #include <linux/io.h> |
| 38 | #include <linux/hardirq.h> |
| 39 | #include <asm/mmu_context.h> |
| 40 | #include <asm/mmu.h> |
| 41 | #include <linux/uaccess.h> |
| 42 | #include <asm/pgtable.h> |
| 43 | #include <asm/cpuinfo.h> |
| 44 | |
| 45 | #ifndef CONFIG_MMU |
| 46 | |
| 47 | /* I have to use dcache values because I can't relate on ram size */ |
| 48 | #define UNCACHED_SHADOW_MASK (cpuinfo.dcache_high - cpuinfo.dcache_base + 1) |
| 49 | |
| 50 | /* |
| 51 | * Consistent memory allocators. Used for DMA devices that want to |
| 52 | * share uncached memory with the processor core. |
| 53 | * My crufty no-MMU approach is simple. In the HW platform we can optionally |
| 54 | * mirror the DDR up above the processor cacheable region. So, memory accessed |
| 55 | * in this mirror region will not be cached. It's alloced from the same |
| 56 | * pool as normal memory, but the handle we return is shifted up into the |
| 57 | * uncached region. This will no doubt cause big problems if memory allocated |
| 58 | * here is not also freed properly. -- JW |
| 59 | */ |
| 60 | void *consistent_alloc(int gfp, size_t size, dma_addr_t *dma_handle) |
| 61 | { |
| 62 | struct page *page, *end, *free; |
| 63 | unsigned long order; |
| 64 | void *ret, *virt; |
| 65 | |
| 66 | if (in_interrupt()) |
| 67 | BUG(); |
| 68 | |
| 69 | size = PAGE_ALIGN(size); |
| 70 | order = get_order(size); |
| 71 | |
| 72 | page = alloc_pages(gfp, order); |
| 73 | if (!page) |
| 74 | goto no_page; |
| 75 | |
| 76 | /* We could do with a page_to_phys and page_to_bus here. */ |
| 77 | virt = page_address(page); |
| 78 | ret = ioremap(virt_to_phys(virt), size); |
| 79 | if (!ret) |
| 80 | goto no_remap; |
| 81 | |
| 82 | /* |
| 83 | * Here's the magic! Note if the uncached shadow is not implemented, |
| 84 | * it's up to the calling code to also test that condition and make |
| 85 | * other arranegments, such as manually flushing the cache and so on. |
| 86 | */ |
| 87 | #ifdef CONFIG_XILINX_UNCACHED_SHADOW |
| 88 | ret = (void *)((unsigned) ret | UNCACHED_SHADOW_MASK); |
| 89 | #endif |
| 90 | /* dma_handle is same as physical (shadowed) address */ |
| 91 | *dma_handle = (dma_addr_t)ret; |
| 92 | |
| 93 | /* |
| 94 | * free wasted pages. We skip the first page since we know |
| 95 | * that it will have count = 1 and won't require freeing. |
| 96 | * We also mark the pages in use as reserved so that |
| 97 | * remap_page_range works. |
| 98 | */ |
| 99 | page = virt_to_page(virt); |
| 100 | free = page + (size >> PAGE_SHIFT); |
| 101 | end = page + (1 << order); |
| 102 | |
| 103 | for (; page < end; page++) { |
| 104 | init_page_count(page); |
| 105 | if (page >= free) |
| 106 | __free_page(page); |
| 107 | else |
| 108 | SetPageReserved(page); |
| 109 | } |
| 110 | |
| 111 | return ret; |
| 112 | no_remap: |
| 113 | __free_pages(page, order); |
| 114 | no_page: |
| 115 | return NULL; |
| 116 | } |
| 117 | |
| 118 | #else |
| 119 | |
| 120 | void *consistent_alloc(int gfp, size_t size, dma_addr_t *dma_handle) |
| 121 | { |
| 122 | int order, err, i; |
| 123 | unsigned long page, va, flags; |
| 124 | phys_addr_t pa; |
| 125 | struct vm_struct *area; |
| 126 | void *ret; |
| 127 | |
| 128 | if (in_interrupt()) |
| 129 | BUG(); |
| 130 | |
| 131 | /* Only allocate page size areas. */ |
| 132 | size = PAGE_ALIGN(size); |
| 133 | order = get_order(size); |
| 134 | |
| 135 | page = __get_free_pages(gfp, order); |
| 136 | if (!page) { |
| 137 | BUG(); |
| 138 | return NULL; |
| 139 | } |
| 140 | |
| 141 | /* |
| 142 | * we need to ensure that there are no cachelines in use, |
| 143 | * or worse dirty in this area. |
| 144 | */ |
| 145 | flush_dcache_range(virt_to_phys(page), virt_to_phys(page) + size); |
| 146 | |
| 147 | /* Allocate some common virtual space to map the new pages. */ |
| 148 | area = get_vm_area(size, VM_ALLOC); |
| 149 | if (area == NULL) { |
| 150 | free_pages(page, order); |
| 151 | return NULL; |
| 152 | } |
| 153 | va = (unsigned long) area->addr; |
| 154 | ret = (void *)va; |
| 155 | |
| 156 | /* This gives us the real physical address of the first page. */ |
| 157 | *dma_handle = pa = virt_to_bus((void *)page); |
| 158 | |
| 159 | /* MS: This is the whole magic - use cache inhibit pages */ |
| 160 | flags = _PAGE_KERNEL | _PAGE_NO_CACHE; |
| 161 | |
| 162 | /* |
| 163 | * Set refcount=1 on all pages in an order>0 |
| 164 | * allocation so that vfree() will actually |
| 165 | * free all pages that were allocated. |
| 166 | */ |
| 167 | if (order > 0) { |
| 168 | struct page *rpage = virt_to_page(page); |
| 169 | for (i = 1; i < (1 << order); i++) |
| 170 | init_page_count(rpage+i); |
| 171 | } |
| 172 | |
| 173 | err = 0; |
| 174 | for (i = 0; i < size && err == 0; i += PAGE_SIZE) |
| 175 | err = map_page(va+i, pa+i, flags); |
| 176 | |
| 177 | if (err) { |
| 178 | vfree((void *)va); |
| 179 | return NULL; |
| 180 | } |
| 181 | |
| 182 | return ret; |
| 183 | } |
| 184 | #endif /* CONFIG_MMU */ |
| 185 | EXPORT_SYMBOL(consistent_alloc); |
| 186 | |
| 187 | /* |
| 188 | * free page(s) as defined by the above mapping. |
| 189 | */ |
| 190 | void consistent_free(void *vaddr) |
| 191 | { |
| 192 | if (in_interrupt()) |
| 193 | BUG(); |
| 194 | |
| 195 | /* Clear SHADOW_MASK bit in address, and free as per usual */ |
| 196 | #ifdef CONFIG_XILINX_UNCACHED_SHADOW |
| 197 | vaddr = (void *)((unsigned)vaddr & ~UNCACHED_SHADOW_MASK); |
| 198 | #endif |
| 199 | vfree(vaddr); |
| 200 | } |
| 201 | EXPORT_SYMBOL(consistent_free); |
| 202 | |
| 203 | /* |
| 204 | * make an area consistent. |
| 205 | */ |
| 206 | void consistent_sync(void *vaddr, size_t size, int direction) |
| 207 | { |
| 208 | unsigned long start; |
| 209 | unsigned long end; |
| 210 | |
| 211 | start = (unsigned long)vaddr; |
| 212 | |
| 213 | /* Convert start address back down to unshadowed memory region */ |
| 214 | #ifdef CONFIG_XILINX_UNCACHED_SHADOW |
| 215 | start &= ~UNCACHED_SHADOW_MASK; |
| 216 | #endif |
| 217 | end = start + size; |
| 218 | |
| 219 | switch (direction) { |
| 220 | case PCI_DMA_NONE: |
| 221 | BUG(); |
| 222 | case PCI_DMA_FROMDEVICE: /* invalidate only */ |
| 223 | flush_dcache_range(start, end); |
| 224 | break; |
| 225 | case PCI_DMA_TODEVICE: /* writeback only */ |
| 226 | flush_dcache_range(start, end); |
| 227 | break; |
| 228 | case PCI_DMA_BIDIRECTIONAL: /* writeback and invalidate */ |
| 229 | flush_dcache_range(start, end); |
| 230 | break; |
| 231 | } |
| 232 | } |
| 233 | EXPORT_SYMBOL(consistent_sync); |
| 234 | |
| 235 | /* |
| 236 | * consistent_sync_page makes memory consistent. identical |
| 237 | * to consistent_sync, but takes a struct page instead of a |
| 238 | * virtual address |
| 239 | */ |
| 240 | void consistent_sync_page(struct page *page, unsigned long offset, |
| 241 | size_t size, int direction) |
| 242 | { |
| 243 | unsigned long start = (unsigned long)page_address(page) + offset; |
| 244 | consistent_sync((void *)start, size, direction); |
| 245 | } |
| 246 | EXPORT_SYMBOL(consistent_sync_page); |