2 * Microblaze support for cache consistent memory.
4 * Copyright (C) 2010 PetaLogix
7 * Based on PowerPC version derived from arch/arm/mm/consistent.c
9 * Copyright (C) 2000 Russell King
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.
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>
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 #include <linux/gfp.h>
37 #include <asm/pgalloc.h>
39 #include <linux/hardirq.h>
40 #include <asm/mmu_context.h>
42 #include <linux/uaccess.h>
43 #include <asm/pgtable.h>
44 #include <asm/cpuinfo.h>
45 #include <asm/tlbflush.h>
48 /* I have to use dcache values because I can't relate on ram size */
49 # define UNCACHED_SHADOW_MASK (cpuinfo.dcache_high - cpuinfo.dcache_base + 1)
53 * Consistent memory allocators. Used for DMA devices that want to
54 * share uncached memory with the processor core.
55 * My crufty no-MMU approach is simple. In the HW platform we can optionally
56 * mirror the DDR up above the processor cacheable region. So, memory accessed
57 * in this mirror region will not be cached. It's alloced from the same
58 * pool as normal memory, but the handle we return is shifted up into the
59 * uncached region. This will no doubt cause big problems if memory allocated
60 * here is not also freed properly. -- JW
62 void *consistent_alloc(gfp_t gfp, size_t size, dma_addr_t *dma_handle)
64 unsigned long order, vaddr;
66 unsigned int i, err = 0;
67 struct page *page, *end;
71 struct vm_struct *area;
78 /* Only allocate page size areas. */
79 size = PAGE_ALIGN(size);
80 order = get_order(size);
82 vaddr = __get_free_pages(gfp, order);
87 * we need to ensure that there are no cachelines in use,
88 * or worse dirty in this area.
90 flush_dcache_range(virt_to_phys((void *)vaddr),
91 virt_to_phys((void *)vaddr) + size);
96 * Here's the magic! Note if the uncached shadow is not implemented,
97 * it's up to the calling code to also test that condition and make
98 * other arranegments, such as manually flushing the cache and so on.
100 # ifdef CONFIG_XILINX_UNCACHED_SHADOW
101 ret = (void *)((unsigned) ret | UNCACHED_SHADOW_MASK);
103 if ((unsigned int)ret > cpuinfo.dcache_base &&
104 (unsigned int)ret < cpuinfo.dcache_high)
106 "ERROR: Your cache coherent area is CACHED!!!\n");
108 /* dma_handle is same as physical (shadowed) address */
109 *dma_handle = (dma_addr_t)ret;
111 /* Allocate some common virtual space to map the new pages. */
112 area = get_vm_area(size, VM_ALLOC);
114 free_pages(vaddr, order);
117 va = (unsigned long) area->addr;
120 /* This gives us the real physical address of the first page. */
121 *dma_handle = pa = virt_to_bus((void *)vaddr);
125 * free wasted pages. We skip the first page since we know
126 * that it will have count = 1 and won't require freeing.
127 * We also mark the pages in use as reserved so that
128 * remap_page_range works.
130 page = virt_to_page(vaddr);
131 end = page + (1 << order);
133 split_page(page, order);
135 for (i = 0; i < size && err == 0; i += PAGE_SIZE) {
137 /* MS: This is the whole magic - use cache inhibit pages */
138 err = map_page(va + i, pa + i, _PAGE_KERNEL | _PAGE_NO_CACHE);
141 SetPageReserved(page);
145 /* Free the otherwise unused pages. */
152 free_pages(vaddr, order);
158 EXPORT_SYMBOL(consistent_alloc);
161 * free page(s) as defined by the above mapping.
163 void consistent_free(size_t size, void *vaddr)
170 size = PAGE_ALIGN(size);
173 /* Clear SHADOW_MASK bit in address, and free as per usual */
174 # ifdef CONFIG_XILINX_UNCACHED_SHADOW
175 vaddr = (void *)((unsigned)vaddr & ~UNCACHED_SHADOW_MASK);
177 page = virt_to_page(vaddr);
180 ClearPageReserved(page);
183 } while (size -= PAGE_SIZE);
189 ptep = pte_offset_kernel(pmd_offset(pgd_offset_k(
190 (unsigned int)vaddr),
191 (unsigned int)vaddr),
192 (unsigned int)vaddr);
193 if (!pte_none(*ptep) && pte_present(*ptep)) {
194 pfn = pte_pfn(*ptep);
195 pte_clear(&init_mm, (unsigned int)vaddr, ptep);
196 if (pfn_valid(pfn)) {
197 page = pfn_to_page(pfn);
199 ClearPageReserved(page);
204 } while (size -= PAGE_SIZE);
210 EXPORT_SYMBOL(consistent_free);
213 * make an area consistent.
215 void consistent_sync(void *vaddr, size_t size, int direction)
220 start = (unsigned long)vaddr;
222 /* Convert start address back down to unshadowed memory region */
223 #ifdef CONFIG_XILINX_UNCACHED_SHADOW
224 start &= ~UNCACHED_SHADOW_MASK;
231 case PCI_DMA_FROMDEVICE: /* invalidate only */
232 invalidate_dcache_range(start, end);
234 case PCI_DMA_TODEVICE: /* writeback only */
235 flush_dcache_range(start, end);
237 case PCI_DMA_BIDIRECTIONAL: /* writeback and invalidate */
238 flush_dcache_range(start, end);
242 EXPORT_SYMBOL(consistent_sync);
245 * consistent_sync_page makes memory consistent. identical
246 * to consistent_sync, but takes a struct page instead of a
249 void consistent_sync_page(struct page *page, unsigned long offset,
250 size_t size, int direction)
252 unsigned long start = (unsigned long)page_address(page) + offset;
253 consistent_sync((void *)start, size, direction);
255 EXPORT_SYMBOL(consistent_sync_page);