Loading...
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
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
18void arch_dma_prep_coherent(struct page *page, size_t size)
19{
20 /*
21 * Evict any existing L1 and/or L2 lines for the backing page
22 * in case it was used earlier as a normal "cached" page.
23 * Yeah this bit us - STAR 9000898266
24 *
25 * Although core does call flush_cache_vmap(), it gets kvaddr hence
26 * can't be used to efficiently flush L1 and/or L2 which need paddr
27 * Currently flush_cache_vmap nukes the L1 cache completely which
28 * will be optimized as a separate commit
29 */
30 dma_cache_wback_inv(page_to_phys(page), size);
31}
32
33/*
34 * Cache operations depending on function and direction argument, inspired by
35 * https://lkml.org/lkml/2018/5/18/979
36 * "dma_sync_*_for_cpu and direction=TO_DEVICE (was Re: [PATCH 02/20]
37 * dma-mapping: provide a generic dma-noncoherent implementation)"
38 *
39 * | map == for_device | unmap == for_cpu
40 * |----------------------------------------------------------------
41 * TO_DEV | writeback writeback | none none
42 * FROM_DEV | invalidate invalidate | invalidate* invalidate*
43 * BIDIR | writeback+inv writeback+inv | invalidate invalidate
44 *
45 * [*] needed for CPU speculative prefetches
46 *
47 * NOTE: we don't check the validity of direction argument as it is done in
48 * upper layer functions (in include/linux/dma-mapping.h)
49 */
50
51void arch_sync_dma_for_device(phys_addr_t paddr, size_t size,
52 enum dma_data_direction dir)
53{
54 switch (dir) {
55 case DMA_TO_DEVICE:
56 dma_cache_wback(paddr, size);
57 break;
58
59 case DMA_FROM_DEVICE:
60 dma_cache_inv(paddr, size);
61 break;
62
63 case DMA_BIDIRECTIONAL:
64 dma_cache_wback_inv(paddr, size);
65 break;
66
67 default:
68 break;
69 }
70}
71
72void arch_sync_dma_for_cpu(phys_addr_t paddr, size_t size,
73 enum dma_data_direction dir)
74{
75 switch (dir) {
76 case DMA_TO_DEVICE:
77 break;
78
79 /* FROM_DEVICE invalidate needed if speculative CPU prefetch only */
80 case DMA_FROM_DEVICE:
81 case DMA_BIDIRECTIONAL:
82 dma_cache_inv(paddr, size);
83 break;
84
85 default:
86 break;
87 }
88}
89
90/*
91 * Plug in direct dma map ops.
92 */
93void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
94 const struct iommu_ops *iommu, bool coherent)
95{
96 /*
97 * IOC hardware snoops all DMA traffic keeping the caches consistent
98 * with memory - eliding need for any explicit cache maintenance of
99 * DMA buffers.
100 */
101 if (is_isa_arcv2() && ioc_enable && coherent)
102 dev->dma_coherent = true;
103
104 dev_info(dev, "use %scoherent DMA ops\n",
105 dev->dma_coherent ? "" : "non");
106}
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-mapping.h>
20#include <asm/cache.h>
21#include <asm/cacheflush.h>
22
23
24static void *arc_dma_alloc(struct device *dev, size_t size,
25 dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
26{
27 unsigned long order = get_order(size);
28 struct page *page;
29 phys_addr_t paddr;
30 void *kvaddr;
31 int need_coh = 1, need_kvaddr = 0;
32
33 page = alloc_pages(gfp, order);
34 if (!page)
35 return NULL;
36
37 /*
38 * IOC relies on all data (even coherent DMA data) being in cache
39 * Thus allocate normal cached memory
40 *
41 * The gains with IOC are two pronged:
42 * -For streaming data, elides need for cache maintenance, saving
43 * cycles in flush code, and bus bandwidth as all the lines of a
44 * buffer need to be flushed out to memory
45 * -For coherent data, Read/Write to buffers terminate early in cache
46 * (vs. always going to memory - thus are faster)
47 */
48 if ((is_isa_arcv2() && ioc_enable) ||
49 (attrs & DMA_ATTR_NON_CONSISTENT))
50 need_coh = 0;
51
52 /*
53 * - A coherent buffer needs MMU mapping to enforce non-cachability
54 * - A highmem page needs a virtual handle (hence MMU mapping)
55 * independent of cachability
56 */
57 if (PageHighMem(page) || need_coh)
58 need_kvaddr = 1;
59
60 /* This is linear addr (0x8000_0000 based) */
61 paddr = page_to_phys(page);
62
63 *dma_handle = paddr;
64
65 /* This is kernel Virtual address (0x7000_0000 based) */
66 if (need_kvaddr) {
67 kvaddr = ioremap_nocache(paddr, size);
68 if (kvaddr == NULL) {
69 __free_pages(page, order);
70 return NULL;
71 }
72 } else {
73 kvaddr = (void *)(u32)paddr;
74 }
75
76 /*
77 * Evict any existing L1 and/or L2 lines for the backing page
78 * in case it was used earlier as a normal "cached" page.
79 * Yeah this bit us - STAR 9000898266
80 *
81 * Although core does call flush_cache_vmap(), it gets kvaddr hence
82 * can't be used to efficiently flush L1 and/or L2 which need paddr
83 * Currently flush_cache_vmap nukes the L1 cache completely which
84 * will be optimized as a separate commit
85 */
86 if (need_coh)
87 dma_cache_wback_inv(paddr, size);
88
89 return kvaddr;
90}
91
92static void arc_dma_free(struct device *dev, size_t size, void *vaddr,
93 dma_addr_t dma_handle, unsigned long attrs)
94{
95 phys_addr_t paddr = dma_handle;
96 struct page *page = virt_to_page(paddr);
97 int is_non_coh = 1;
98
99 is_non_coh = (attrs & DMA_ATTR_NON_CONSISTENT) ||
100 (is_isa_arcv2() && ioc_enable);
101
102 if (PageHighMem(page) || !is_non_coh)
103 iounmap((void __force __iomem *)vaddr);
104
105 __free_pages(page, get_order(size));
106}
107
108static int arc_dma_mmap(struct device *dev, struct vm_area_struct *vma,
109 void *cpu_addr, dma_addr_t dma_addr, size_t size,
110 unsigned long attrs)
111{
112 unsigned long user_count = vma_pages(vma);
113 unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT;
114 unsigned long pfn = __phys_to_pfn(dma_addr);
115 unsigned long off = vma->vm_pgoff;
116 int ret = -ENXIO;
117
118 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
119
120 if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
121 return ret;
122
123 if (off < count && user_count <= (count - off)) {
124 ret = remap_pfn_range(vma, vma->vm_start,
125 pfn + off,
126 user_count << PAGE_SHIFT,
127 vma->vm_page_prot);
128 }
129
130 return ret;
131}
132
133/*
134 * streaming DMA Mapping API...
135 * CPU accesses page via normal paddr, thus needs to explicitly made
136 * consistent before each use
137 */
138static void _dma_cache_sync(phys_addr_t paddr, size_t size,
139 enum dma_data_direction dir)
140{
141 switch (dir) {
142 case DMA_FROM_DEVICE:
143 dma_cache_inv(paddr, size);
144 break;
145 case DMA_TO_DEVICE:
146 dma_cache_wback(paddr, size);
147 break;
148 case DMA_BIDIRECTIONAL:
149 dma_cache_wback_inv(paddr, size);
150 break;
151 default:
152 pr_err("Invalid DMA dir [%d] for OP @ %pa[p]\n", dir, &paddr);
153 }
154}
155
156/*
157 * arc_dma_map_page - map a portion of a page for streaming DMA
158 *
159 * Ensure that any data held in the cache is appropriately discarded
160 * or written back.
161 *
162 * The device owns this memory once this call has completed. The CPU
163 * can regain ownership by calling dma_unmap_page().
164 *
165 * Note: while it takes struct page as arg, caller can "abuse" it to pass
166 * a region larger than PAGE_SIZE, provided it is physically contiguous
167 * and this still works correctly
168 */
169static dma_addr_t arc_dma_map_page(struct device *dev, struct page *page,
170 unsigned long offset, size_t size, enum dma_data_direction dir,
171 unsigned long attrs)
172{
173 phys_addr_t paddr = page_to_phys(page) + offset;
174
175 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
176 _dma_cache_sync(paddr, size, dir);
177
178 return paddr;
179}
180
181/*
182 * arc_dma_unmap_page - unmap a buffer previously mapped through dma_map_page()
183 *
184 * After this call, reads by the CPU to the buffer are guaranteed to see
185 * whatever the device wrote there.
186 *
187 * Note: historically this routine was not implemented for ARC
188 */
189static void arc_dma_unmap_page(struct device *dev, dma_addr_t handle,
190 size_t size, enum dma_data_direction dir,
191 unsigned long attrs)
192{
193 phys_addr_t paddr = handle;
194
195 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
196 _dma_cache_sync(paddr, size, dir);
197}
198
199static int arc_dma_map_sg(struct device *dev, struct scatterlist *sg,
200 int nents, enum dma_data_direction dir, unsigned long attrs)
201{
202 struct scatterlist *s;
203 int i;
204
205 for_each_sg(sg, s, nents, i)
206 s->dma_address = dma_map_page(dev, sg_page(s), s->offset,
207 s->length, dir);
208
209 return nents;
210}
211
212static void arc_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
213 int nents, enum dma_data_direction dir,
214 unsigned long attrs)
215{
216 struct scatterlist *s;
217 int i;
218
219 for_each_sg(sg, s, nents, i)
220 arc_dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir,
221 attrs);
222}
223
224static void arc_dma_sync_single_for_cpu(struct device *dev,
225 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
226{
227 _dma_cache_sync(dma_handle, size, DMA_FROM_DEVICE);
228}
229
230static void arc_dma_sync_single_for_device(struct device *dev,
231 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
232{
233 _dma_cache_sync(dma_handle, size, DMA_TO_DEVICE);
234}
235
236static void arc_dma_sync_sg_for_cpu(struct device *dev,
237 struct scatterlist *sglist, int nelems,
238 enum dma_data_direction dir)
239{
240 int i;
241 struct scatterlist *sg;
242
243 for_each_sg(sglist, sg, nelems, i)
244 _dma_cache_sync(sg_phys(sg), sg->length, dir);
245}
246
247static void arc_dma_sync_sg_for_device(struct device *dev,
248 struct scatterlist *sglist, int nelems,
249 enum dma_data_direction dir)
250{
251 int i;
252 struct scatterlist *sg;
253
254 for_each_sg(sglist, sg, nelems, i)
255 _dma_cache_sync(sg_phys(sg), sg->length, dir);
256}
257
258static int arc_dma_supported(struct device *dev, u64 dma_mask)
259{
260 /* Support 32 bit DMA mask exclusively */
261 return dma_mask == DMA_BIT_MASK(32);
262}
263
264const struct dma_map_ops arc_dma_ops = {
265 .alloc = arc_dma_alloc,
266 .free = arc_dma_free,
267 .mmap = arc_dma_mmap,
268 .map_page = arc_dma_map_page,
269 .unmap_page = arc_dma_unmap_page,
270 .map_sg = arc_dma_map_sg,
271 .unmap_sg = arc_dma_unmap_sg,
272 .sync_single_for_device = arc_dma_sync_single_for_device,
273 .sync_single_for_cpu = arc_dma_sync_single_for_cpu,
274 .sync_sg_for_cpu = arc_dma_sync_sg_for_cpu,
275 .sync_sg_for_device = arc_dma_sync_sg_for_device,
276 .dma_supported = arc_dma_supported,
277};
278EXPORT_SYMBOL(arc_dma_ops);