xref: /openbmc/linux/arch/arc/mm/dma.c (revision 3d92aa45)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
4  */
5 
6 #include <linux/dma-noncoherent.h>
7 #include <asm/cache.h>
8 #include <asm/cacheflush.h>
9 
10 /*
11  * ARCH specific callbacks for generic noncoherent DMA ops (dma/noncoherent.c)
12  *  - hardware IOC not available (or "dma-coherent" not set for device in DT)
13  *  - But still handle both coherent and non-coherent requests from caller
14  *
15  * For DMA coherent hardware (IOC) generic code suffices
16  */
17 void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
18 		gfp_t gfp, unsigned long attrs)
19 {
20 	unsigned long order = get_order(size);
21 	struct page *page;
22 	phys_addr_t paddr;
23 	void *kvaddr;
24 	bool need_coh = !(attrs & DMA_ATTR_NON_CONSISTENT);
25 
26 	/*
27 	 * __GFP_HIGHMEM flag is cleared by upper layer functions
28 	 * (in include/linux/dma-mapping.h) so we should never get a
29 	 * __GFP_HIGHMEM here.
30 	 */
31 	BUG_ON(gfp & __GFP_HIGHMEM);
32 
33 	page = alloc_pages(gfp | __GFP_ZERO, order);
34 	if (!page)
35 		return NULL;
36 
37 	/* This is linear addr (0x8000_0000 based) */
38 	paddr = page_to_phys(page);
39 
40 	*dma_handle = paddr;
41 
42 	/*
43 	 * A coherent buffer needs MMU mapping to enforce non-cachability.
44 	 * kvaddr is kernel Virtual address (0x7000_0000 based).
45 	 */
46 	if (need_coh) {
47 		kvaddr = ioremap_nocache(paddr, size);
48 		if (kvaddr == NULL) {
49 			__free_pages(page, order);
50 			return NULL;
51 		}
52 	} else {
53 		kvaddr = (void *)(u32)paddr;
54 	}
55 
56 	/*
57 	 * Evict any existing L1 and/or L2 lines for the backing page
58 	 * in case it was used earlier as a normal "cached" page.
59 	 * Yeah this bit us - STAR 9000898266
60 	 *
61 	 * Although core does call flush_cache_vmap(), it gets kvaddr hence
62 	 * can't be used to efficiently flush L1 and/or L2 which need paddr
63 	 * Currently flush_cache_vmap nukes the L1 cache completely which
64 	 * will be optimized as a separate commit
65 	 */
66 	if (need_coh)
67 		dma_cache_wback_inv(paddr, size);
68 
69 	return kvaddr;
70 }
71 
72 void arch_dma_free(struct device *dev, size_t size, void *vaddr,
73 		dma_addr_t dma_handle, unsigned long attrs)
74 {
75 	phys_addr_t paddr = dma_handle;
76 	struct page *page = virt_to_page(paddr);
77 
78 	if (!(attrs & DMA_ATTR_NON_CONSISTENT))
79 		iounmap((void __force __iomem *)vaddr);
80 
81 	__free_pages(page, get_order(size));
82 }
83 
84 long arch_dma_coherent_to_pfn(struct device *dev, void *cpu_addr,
85 		dma_addr_t dma_addr)
86 {
87 	return __phys_to_pfn(dma_addr);
88 }
89 
90 /*
91  * Cache operations depending on function and direction argument, inspired by
92  * https://lkml.org/lkml/2018/5/18/979
93  * "dma_sync_*_for_cpu and direction=TO_DEVICE (was Re: [PATCH 02/20]
94  * dma-mapping: provide a generic dma-noncoherent implementation)"
95  *
96  *          |   map          ==  for_device     |   unmap     ==  for_cpu
97  *          |----------------------------------------------------------------
98  * TO_DEV   |   writeback        writeback      |   none          none
99  * FROM_DEV |   invalidate       invalidate     |   invalidate*   invalidate*
100  * BIDIR    |   writeback+inv    writeback+inv  |   invalidate    invalidate
101  *
102  *     [*] needed for CPU speculative prefetches
103  *
104  * NOTE: we don't check the validity of direction argument as it is done in
105  * upper layer functions (in include/linux/dma-mapping.h)
106  */
107 
108 void arch_sync_dma_for_device(struct device *dev, phys_addr_t paddr,
109 		size_t size, enum dma_data_direction dir)
110 {
111 	switch (dir) {
112 	case DMA_TO_DEVICE:
113 		dma_cache_wback(paddr, size);
114 		break;
115 
116 	case DMA_FROM_DEVICE:
117 		dma_cache_inv(paddr, size);
118 		break;
119 
120 	case DMA_BIDIRECTIONAL:
121 		dma_cache_wback_inv(paddr, size);
122 		break;
123 
124 	default:
125 		break;
126 	}
127 }
128 
129 void arch_sync_dma_for_cpu(struct device *dev, phys_addr_t paddr,
130 		size_t size, enum dma_data_direction dir)
131 {
132 	switch (dir) {
133 	case DMA_TO_DEVICE:
134 		break;
135 
136 	/* FROM_DEVICE invalidate needed if speculative CPU prefetch only */
137 	case DMA_FROM_DEVICE:
138 	case DMA_BIDIRECTIONAL:
139 		dma_cache_inv(paddr, size);
140 		break;
141 
142 	default:
143 		break;
144 	}
145 }
146 
147 /*
148  * Plug in direct dma map ops.
149  */
150 void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
151 			const struct iommu_ops *iommu, bool coherent)
152 {
153 	/*
154 	 * IOC hardware snoops all DMA traffic keeping the caches consistent
155 	 * with memory - eliding need for any explicit cache maintenance of
156 	 * DMA buffers.
157 	 */
158 	if (is_isa_arcv2() && ioc_enable && coherent)
159 		dev->dma_coherent = true;
160 
161 	dev_info(dev, "use %sncoherent DMA ops\n",
162 		 dev->dma_coherent ? "" : "non");
163 }
164