xref: /openbmc/linux/arch/mips/mm/dma-noncoherent.c (revision cce8e04c)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2000  Ani Joshi <ajoshi@unixbox.com>
4  * Copyright (C) 2000, 2001, 06	 Ralf Baechle <ralf@linux-mips.org>
5  * swiped from i386, and cloned for MIPS by Geert, polished by Ralf.
6  */
7 #include <linux/dma-direct.h>
8 #include <linux/dma-noncoherent.h>
9 #include <linux/dma-contiguous.h>
10 #include <linux/highmem.h>
11 
12 #include <asm/cache.h>
13 #include <asm/cpu-type.h>
14 #include <asm/dma-coherence.h>
15 #include <asm/io.h>
16 
17 /*
18  * The affected CPUs below in 'cpu_needs_post_dma_flush()' can speculatively
19  * fill random cachelines with stale data at any time, requiring an extra
20  * flush post-DMA.
21  *
22  * Warning on the terminology - Linux calls an uncached area coherent;  MIPS
23  * terminology calls memory areas with hardware maintained coherency coherent.
24  *
25  * Note that the R14000 and R16000 should also be checked for in this condition.
26  * However this function is only called on non-I/O-coherent systems and only the
27  * R10000 and R12000 are used in such systems, the SGI IP28 Indigo² rsp.
28  * SGI IP32 aka O2.
29  */
30 static inline bool cpu_needs_post_dma_flush(struct device *dev)
31 {
32 	switch (boot_cpu_type()) {
33 	case CPU_R10000:
34 	case CPU_R12000:
35 	case CPU_BMIPS5000:
36 		return true;
37 	default:
38 		/*
39 		 * Presence of MAARs suggests that the CPU supports
40 		 * speculatively prefetching data, and therefore requires
41 		 * the post-DMA flush/invalidate.
42 		 */
43 		return cpu_has_maar;
44 	}
45 }
46 
47 void *arch_dma_alloc(struct device *dev, size_t size,
48 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
49 {
50 	void *ret;
51 
52 	ret = dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
53 	if (ret && !(attrs & DMA_ATTR_NON_CONSISTENT)) {
54 		dma_cache_wback_inv((unsigned long) ret, size);
55 		ret = (void *)UNCAC_ADDR(ret);
56 	}
57 
58 	return ret;
59 }
60 
61 void arch_dma_free(struct device *dev, size_t size, void *cpu_addr,
62 		dma_addr_t dma_addr, unsigned long attrs)
63 {
64 	if (!(attrs & DMA_ATTR_NON_CONSISTENT))
65 		cpu_addr = (void *)CAC_ADDR((unsigned long)cpu_addr);
66 	dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
67 }
68 
69 long arch_dma_coherent_to_pfn(struct device *dev, void *cpu_addr,
70 		dma_addr_t dma_addr)
71 {
72 	unsigned long addr = CAC_ADDR((unsigned long)cpu_addr);
73 	return page_to_pfn(virt_to_page((void *)addr));
74 }
75 
76 pgprot_t arch_dma_mmap_pgprot(struct device *dev, pgprot_t prot,
77 		unsigned long attrs)
78 {
79 	if (attrs & DMA_ATTR_WRITE_COMBINE)
80 		return pgprot_writecombine(prot);
81 	return pgprot_noncached(prot);
82 }
83 
84 static inline void dma_sync_virt(void *addr, size_t size,
85 		enum dma_data_direction dir)
86 {
87 	switch (dir) {
88 	case DMA_TO_DEVICE:
89 		dma_cache_wback((unsigned long)addr, size);
90 		break;
91 
92 	case DMA_FROM_DEVICE:
93 		dma_cache_inv((unsigned long)addr, size);
94 		break;
95 
96 	case DMA_BIDIRECTIONAL:
97 		dma_cache_wback_inv((unsigned long)addr, size);
98 		break;
99 
100 	default:
101 		BUG();
102 	}
103 }
104 
105 /*
106  * A single sg entry may refer to multiple physically contiguous pages.  But
107  * we still need to process highmem pages individually.  If highmem is not
108  * configured then the bulk of this loop gets optimized out.
109  */
110 static inline void dma_sync_phys(phys_addr_t paddr, size_t size,
111 		enum dma_data_direction dir)
112 {
113 	struct page *page = pfn_to_page(paddr >> PAGE_SHIFT);
114 	unsigned long offset = paddr & ~PAGE_MASK;
115 	size_t left = size;
116 
117 	do {
118 		size_t len = left;
119 
120 		if (PageHighMem(page)) {
121 			void *addr;
122 
123 			if (offset + len > PAGE_SIZE) {
124 				if (offset >= PAGE_SIZE) {
125 					page += offset >> PAGE_SHIFT;
126 					offset &= ~PAGE_MASK;
127 				}
128 				len = PAGE_SIZE - offset;
129 			}
130 
131 			addr = kmap_atomic(page);
132 			dma_sync_virt(addr + offset, len, dir);
133 			kunmap_atomic(addr);
134 		} else
135 			dma_sync_virt(page_address(page) + offset, size, dir);
136 		offset = 0;
137 		page++;
138 		left -= len;
139 	} while (left);
140 }
141 
142 void arch_sync_dma_for_device(struct device *dev, phys_addr_t paddr,
143 		size_t size, enum dma_data_direction dir)
144 {
145 	dma_sync_phys(paddr, size, dir);
146 }
147 
148 void arch_sync_dma_for_cpu(struct device *dev, phys_addr_t paddr,
149 		size_t size, enum dma_data_direction dir)
150 {
151 	if (cpu_needs_post_dma_flush(dev))
152 		dma_sync_phys(paddr, size, dir);
153 }
154 
155 void arch_dma_cache_sync(struct device *dev, void *vaddr, size_t size,
156 		enum dma_data_direction direction)
157 {
158 	BUG_ON(direction == DMA_NONE);
159 
160 	dma_sync_virt(vaddr, size, direction);
161 }
162