xref: /openbmc/linux/arch/arc/mm/dma.c (revision 5b394b2d)
1 /*
2  * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of the GNU General Public License version 2 as
6  * published by the Free Software Foundation.
7  */
8 
9 /*
10  * DMA Coherent API Notes
11  *
12  * I/O is inherently non-coherent on ARC. So a coherent DMA buffer is
13  * implemented by accessing it using a kernel virtual address, with
14  * Cache bit off in the TLB entry.
15  *
16  * The default DMA address == Phy address which is 0x8000_0000 based.
17  */
18 
19 #include <linux/dma-noncoherent.h>
20 #include <asm/cache.h>
21 #include <asm/cacheflush.h>
22 
23 void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
24 		gfp_t gfp, unsigned long attrs)
25 {
26 	unsigned long order = get_order(size);
27 	struct page *page;
28 	phys_addr_t paddr;
29 	void *kvaddr;
30 	int need_coh = 1, need_kvaddr = 0;
31 
32 	page = alloc_pages(gfp, order);
33 	if (!page)
34 		return NULL;
35 
36 	/*
37 	 * IOC relies on all data (even coherent DMA data) being in cache
38 	 * Thus allocate normal cached memory
39 	 *
40 	 * The gains with IOC are two pronged:
41 	 *   -For streaming data, elides need for cache maintenance, saving
42 	 *    cycles in flush code, and bus bandwidth as all the lines of a
43 	 *    buffer need to be flushed out to memory
44 	 *   -For coherent data, Read/Write to buffers terminate early in cache
45 	 *   (vs. always going to memory - thus are faster)
46 	 */
47 	if ((is_isa_arcv2() && ioc_enable) ||
48 	    (attrs & DMA_ATTR_NON_CONSISTENT))
49 		need_coh = 0;
50 
51 	/*
52 	 * - A coherent buffer needs MMU mapping to enforce non-cachability
53 	 * - A highmem page needs a virtual handle (hence MMU mapping)
54 	 *   independent of cachability
55 	 */
56 	if (PageHighMem(page) || need_coh)
57 		need_kvaddr = 1;
58 
59 	/* This is linear addr (0x8000_0000 based) */
60 	paddr = page_to_phys(page);
61 
62 	*dma_handle = paddr;
63 
64 	/* This is kernel Virtual address (0x7000_0000 based) */
65 	if (need_kvaddr) {
66 		kvaddr = ioremap_nocache(paddr, size);
67 		if (kvaddr == NULL) {
68 			__free_pages(page, order);
69 			return NULL;
70 		}
71 	} else {
72 		kvaddr = (void *)(u32)paddr;
73 	}
74 
75 	/*
76 	 * Evict any existing L1 and/or L2 lines for the backing page
77 	 * in case it was used earlier as a normal "cached" page.
78 	 * Yeah this bit us - STAR 9000898266
79 	 *
80 	 * Although core does call flush_cache_vmap(), it gets kvaddr hence
81 	 * can't be used to efficiently flush L1 and/or L2 which need paddr
82 	 * Currently flush_cache_vmap nukes the L1 cache completely which
83 	 * will be optimized as a separate commit
84 	 */
85 	if (need_coh)
86 		dma_cache_wback_inv(paddr, size);
87 
88 	return kvaddr;
89 }
90 
91 void arch_dma_free(struct device *dev, size_t size, void *vaddr,
92 		dma_addr_t dma_handle, unsigned long attrs)
93 {
94 	phys_addr_t paddr = dma_handle;
95 	struct page *page = virt_to_page(paddr);
96 	int is_non_coh = 1;
97 
98 	is_non_coh = (attrs & DMA_ATTR_NON_CONSISTENT) ||
99 			(is_isa_arcv2() && ioc_enable);
100 
101 	if (PageHighMem(page) || !is_non_coh)
102 		iounmap((void __force __iomem *)vaddr);
103 
104 	__free_pages(page, get_order(size));
105 }
106 
107 int arch_dma_mmap(struct device *dev, struct vm_area_struct *vma,
108 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
109 		unsigned long attrs)
110 {
111 	unsigned long user_count = vma_pages(vma);
112 	unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT;
113 	unsigned long pfn = __phys_to_pfn(dma_addr);
114 	unsigned long off = vma->vm_pgoff;
115 	int ret = -ENXIO;
116 
117 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
118 
119 	if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
120 		return ret;
121 
122 	if (off < count && user_count <= (count - off)) {
123 		ret = remap_pfn_range(vma, vma->vm_start,
124 				      pfn + off,
125 				      user_count << PAGE_SHIFT,
126 				      vma->vm_page_prot);
127 	}
128 
129 	return ret;
130 }
131 
132 /*
133  * Cache operations depending on function and direction argument, inspired by
134  * https://lkml.org/lkml/2018/5/18/979
135  * "dma_sync_*_for_cpu and direction=TO_DEVICE (was Re: [PATCH 02/20]
136  * dma-mapping: provide a generic dma-noncoherent implementation)"
137  *
138  *          |   map          ==  for_device     |   unmap     ==  for_cpu
139  *          |----------------------------------------------------------------
140  * TO_DEV   |   writeback        writeback      |   none          none
141  * FROM_DEV |   invalidate       invalidate     |   invalidate*   invalidate*
142  * BIDIR    |   writeback+inv    writeback+inv  |   invalidate    invalidate
143  *
144  *     [*] needed for CPU speculative prefetches
145  *
146  * NOTE: we don't check the validity of direction argument as it is done in
147  * upper layer functions (in include/linux/dma-mapping.h)
148  */
149 
150 void arch_sync_dma_for_device(struct device *dev, phys_addr_t paddr,
151 		size_t size, enum dma_data_direction dir)
152 {
153 	switch (dir) {
154 	case DMA_TO_DEVICE:
155 		dma_cache_wback(paddr, size);
156 		break;
157 
158 	case DMA_FROM_DEVICE:
159 		dma_cache_inv(paddr, size);
160 		break;
161 
162 	case DMA_BIDIRECTIONAL:
163 		dma_cache_wback_inv(paddr, size);
164 		break;
165 
166 	default:
167 		break;
168 	}
169 }
170 
171 void arch_sync_dma_for_cpu(struct device *dev, phys_addr_t paddr,
172 		size_t size, enum dma_data_direction dir)
173 {
174 	switch (dir) {
175 	case DMA_TO_DEVICE:
176 		break;
177 
178 	/* FROM_DEVICE invalidate needed if speculative CPU prefetch only */
179 	case DMA_FROM_DEVICE:
180 	case DMA_BIDIRECTIONAL:
181 		dma_cache_inv(paddr, size);
182 		break;
183 
184 	default:
185 		break;
186 	}
187 }
188