xref: /openbmc/linux/arch/openrisc/kernel/dma.c (revision 74ba9207)
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-noncoherent.h>
23 
24 #include <asm/cpuinfo.h>
25 #include <asm/spr_defs.h>
26 #include <asm/tlbflush.h>
27 
28 static int
29 page_set_nocache(pte_t *pte, unsigned long addr,
30 		 unsigned long next, struct mm_walk *walk)
31 {
32 	unsigned long cl;
33 	struct cpuinfo_or1k *cpuinfo = &cpuinfo_or1k[smp_processor_id()];
34 
35 	pte_val(*pte) |= _PAGE_CI;
36 
37 	/*
38 	 * Flush the page out of the TLB so that the new page flags get
39 	 * picked up next time there's an access
40 	 */
41 	flush_tlb_page(NULL, addr);
42 
43 	/* Flush page out of dcache */
44 	for (cl = __pa(addr); cl < __pa(next); cl += cpuinfo->dcache_block_size)
45 		mtspr(SPR_DCBFR, cl);
46 
47 	return 0;
48 }
49 
50 static int
51 page_clear_nocache(pte_t *pte, unsigned long addr,
52 		   unsigned long next, struct mm_walk *walk)
53 {
54 	pte_val(*pte) &= ~_PAGE_CI;
55 
56 	/*
57 	 * Flush the page out of the TLB so that the new page flags get
58 	 * picked up next time there's an access
59 	 */
60 	flush_tlb_page(NULL, addr);
61 
62 	return 0;
63 }
64 
65 /*
66  * Alloc "coherent" memory, which for OpenRISC means simply uncached.
67  *
68  * This function effectively just calls __get_free_pages, sets the
69  * cache-inhibit bit on those pages, and makes sure that the pages are
70  * flushed out of the cache before they are used.
71  *
72  * If the NON_CONSISTENT attribute is set, then this function just
73  * returns "normal", cachable memory.
74  *
75  * There are additional flags WEAK_ORDERING and WRITE_COMBINE to take
76  * into consideration here, too.  All current known implementations of
77  * the OR1K support only strongly ordered memory accesses, so that flag
78  * is being ignored for now; uncached but write-combined memory is a
79  * missing feature of the OR1K.
80  */
81 void *
82 arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
83 		gfp_t gfp, unsigned long attrs)
84 {
85 	unsigned long va;
86 	void *page;
87 	struct mm_walk walk = {
88 		.pte_entry = page_set_nocache,
89 		.mm = &init_mm
90 	};
91 
92 	page = alloc_pages_exact(size, gfp | __GFP_ZERO);
93 	if (!page)
94 		return NULL;
95 
96 	/* This gives us the real physical address of the first page. */
97 	*dma_handle = __pa(page);
98 
99 	va = (unsigned long)page;
100 
101 	if ((attrs & DMA_ATTR_NON_CONSISTENT) == 0) {
102 		/*
103 		 * We need to iterate through the pages, clearing the dcache for
104 		 * them and setting the cache-inhibit bit.
105 		 */
106 		if (walk_page_range(va, va + size, &walk)) {
107 			free_pages_exact(page, size);
108 			return NULL;
109 		}
110 	}
111 
112 	return (void *)va;
113 }
114 
115 void
116 arch_dma_free(struct device *dev, size_t size, void *vaddr,
117 		dma_addr_t dma_handle, unsigned long attrs)
118 {
119 	unsigned long va = (unsigned long)vaddr;
120 	struct mm_walk walk = {
121 		.pte_entry = page_clear_nocache,
122 		.mm = &init_mm
123 	};
124 
125 	if ((attrs & DMA_ATTR_NON_CONSISTENT) == 0) {
126 		/* walk_page_range shouldn't be able to fail here */
127 		WARN_ON(walk_page_range(va, va + size, &walk));
128 	}
129 
130 	free_pages_exact(vaddr, size);
131 }
132 
133 void arch_sync_dma_for_device(struct device *dev, phys_addr_t addr, size_t size,
134 		enum dma_data_direction dir)
135 {
136 	unsigned long cl;
137 	struct cpuinfo_or1k *cpuinfo = &cpuinfo_or1k[smp_processor_id()];
138 
139 	switch (dir) {
140 	case DMA_TO_DEVICE:
141 		/* Flush the dcache for the requested range */
142 		for (cl = addr; cl < addr + size;
143 		     cl += cpuinfo->dcache_block_size)
144 			mtspr(SPR_DCBFR, cl);
145 		break;
146 	case DMA_FROM_DEVICE:
147 		/* Invalidate the dcache for the requested range */
148 		for (cl = addr; cl < addr + size;
149 		     cl += cpuinfo->dcache_block_size)
150 			mtspr(SPR_DCBIR, cl);
151 		break;
152 	default:
153 		/*
154 		 * NOTE: If dir == DMA_BIDIRECTIONAL then there's no need to
155 		 * flush nor invalidate the cache here as the area will need
156 		 * to be manually synced anyway.
157 		 */
158 		break;
159 	}
160 }
161