1 /*
2  * This file is subject to the terms and conditions of the GNU General Public
3  * License.  See the file "COPYING" in the main directory of this archive
4  * for more details.
5  *
6  * Copyright (C) 2000  Ani Joshi <ajoshi@unixbox.com>
7  * Copyright (C) 2000, 2001, 06	 Ralf Baechle <ralf@linux-mips.org>
8  * swiped from i386, and cloned for MIPS by Geert, polished by Ralf.
9  */
10 
11 #include <linux/types.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/mm.h>
14 #include <linux/module.h>
15 #include <linux/scatterlist.h>
16 #include <linux/string.h>
17 #include <linux/gfp.h>
18 #include <linux/highmem.h>
19 #include <linux/dma-contiguous.h>
20 
21 #include <asm/cache.h>
22 #include <asm/cpu-type.h>
23 #include <asm/io.h>
24 
25 #include <dma-coherence.h>
26 
27 #ifdef CONFIG_DMA_MAYBE_COHERENT
28 int coherentio = 0;	/* User defined DMA coherency from command line. */
29 EXPORT_SYMBOL_GPL(coherentio);
30 int hw_coherentio = 0;	/* Actual hardware supported DMA coherency setting. */
31 
setcoherentio(char * str)32 static int __init setcoherentio(char *str)
33 {
34 	coherentio = 1;
35 	pr_info("Hardware DMA cache coherency (command line)\n");
36 	return 0;
37 }
38 early_param("coherentio", setcoherentio);
39 
setnocoherentio(char * str)40 static int __init setnocoherentio(char *str)
41 {
42 	coherentio = 0;
43 	pr_info("Software DMA cache coherency (command line)\n");
44 	return 0;
45 }
46 early_param("nocoherentio", setnocoherentio);
47 #endif
48 
dma_addr_to_page(struct device * dev,dma_addr_t dma_addr)49 static inline struct page *dma_addr_to_page(struct device *dev,
50 	dma_addr_t dma_addr)
51 {
52 	return pfn_to_page(
53 		plat_dma_addr_to_phys(dev, dma_addr) >> PAGE_SHIFT);
54 }
55 
56 /*
57  * The affected CPUs below in 'cpu_needs_post_dma_flush()' can
58  * speculatively fill random cachelines with stale data at any time,
59  * requiring an extra flush post-DMA.
60  *
61  * Warning on the terminology - Linux calls an uncached area coherent;
62  * MIPS terminology calls memory areas with hardware maintained coherency
63  * coherent.
64  *
65  * Note that the R14000 and R16000 should also be checked for in this
66  * condition.  However this function is only called on non-I/O-coherent
67  * systems and only the R10000 and R12000 are used in such systems, the
68  * SGI IP28 Indigo² rsp. SGI IP32 aka O2.
69  */
cpu_needs_post_dma_flush(struct device * dev)70 static inline int cpu_needs_post_dma_flush(struct device *dev)
71 {
72 	return !plat_device_is_coherent(dev) &&
73 	       (boot_cpu_type() == CPU_R10000 ||
74 		boot_cpu_type() == CPU_R12000 ||
75 		boot_cpu_type() == CPU_BMIPS5000);
76 }
77 
massage_gfp_flags(const struct device * dev,gfp_t gfp)78 static gfp_t massage_gfp_flags(const struct device *dev, gfp_t gfp)
79 {
80 	gfp_t dma_flag;
81 
82 	/* ignore region specifiers */
83 	gfp &= ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM);
84 
85 #ifdef CONFIG_ISA
86 	if (dev == NULL)
87 		dma_flag = __GFP_DMA;
88 	else
89 #endif
90 #if defined(CONFIG_ZONE_DMA32) && defined(CONFIG_ZONE_DMA)
91 	     if (dev->coherent_dma_mask < DMA_BIT_MASK(32))
92 			dma_flag = __GFP_DMA;
93 	else if (dev->coherent_dma_mask < DMA_BIT_MASK(64))
94 			dma_flag = __GFP_DMA32;
95 	else
96 #endif
97 #if defined(CONFIG_ZONE_DMA32) && !defined(CONFIG_ZONE_DMA)
98 	     if (dev->coherent_dma_mask < DMA_BIT_MASK(64))
99 		dma_flag = __GFP_DMA32;
100 	else
101 #endif
102 #if defined(CONFIG_ZONE_DMA) && !defined(CONFIG_ZONE_DMA32)
103 	     if (dev->coherent_dma_mask < DMA_BIT_MASK(sizeof(phys_addr_t) * 8))
104 		dma_flag = __GFP_DMA;
105 	else
106 #endif
107 		dma_flag = 0;
108 
109 	/* Don't invoke OOM killer */
110 	gfp |= __GFP_NORETRY;
111 
112 	return gfp | dma_flag;
113 }
114 
dma_alloc_noncoherent(struct device * dev,size_t size,dma_addr_t * dma_handle,gfp_t gfp)115 void *dma_alloc_noncoherent(struct device *dev, size_t size,
116 	dma_addr_t * dma_handle, gfp_t gfp)
117 {
118 	void *ret;
119 
120 	gfp = massage_gfp_flags(dev, gfp);
121 
122 	ret = (void *) __get_free_pages(gfp, get_order(size));
123 
124 	if (ret != NULL) {
125 		memset(ret, 0, size);
126 		*dma_handle = plat_map_dma_mem(dev, ret, size);
127 	}
128 
129 	return ret;
130 }
131 EXPORT_SYMBOL(dma_alloc_noncoherent);
132 
mips_dma_alloc_coherent(struct device * dev,size_t size,dma_addr_t * dma_handle,gfp_t gfp,struct dma_attrs * attrs)133 static void *mips_dma_alloc_coherent(struct device *dev, size_t size,
134 	dma_addr_t * dma_handle, gfp_t gfp, struct dma_attrs *attrs)
135 {
136 	void *ret;
137 	struct page *page = NULL;
138 	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
139 
140 	if (dma_alloc_from_coherent(dev, size, dma_handle, &ret))
141 		return ret;
142 
143 	gfp = massage_gfp_flags(dev, gfp);
144 
145 	if (IS_ENABLED(CONFIG_DMA_CMA) && !(gfp & GFP_ATOMIC))
146 		page = dma_alloc_from_contiguous(dev,
147 					count, get_order(size));
148 	if (!page)
149 		page = alloc_pages(gfp, get_order(size));
150 
151 	if (!page)
152 		return NULL;
153 
154 	ret = page_address(page);
155 	memset(ret, 0, size);
156 	*dma_handle = plat_map_dma_mem(dev, ret, size);
157 	if (!plat_device_is_coherent(dev)) {
158 		dma_cache_wback_inv((unsigned long) ret, size);
159 		if (!hw_coherentio)
160 			ret = UNCAC_ADDR(ret);
161 	}
162 
163 	return ret;
164 }
165 
166 
dma_free_noncoherent(struct device * dev,size_t size,void * vaddr,dma_addr_t dma_handle)167 void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr,
168 	dma_addr_t dma_handle)
169 {
170 	plat_unmap_dma_mem(dev, dma_handle, size, DMA_BIDIRECTIONAL);
171 	free_pages((unsigned long) vaddr, get_order(size));
172 }
173 EXPORT_SYMBOL(dma_free_noncoherent);
174 
mips_dma_free_coherent(struct device * dev,size_t size,void * vaddr,dma_addr_t dma_handle,struct dma_attrs * attrs)175 static void mips_dma_free_coherent(struct device *dev, size_t size, void *vaddr,
176 	dma_addr_t dma_handle, struct dma_attrs *attrs)
177 {
178 	unsigned long addr = (unsigned long) vaddr;
179 	int order = get_order(size);
180 	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
181 	struct page *page = NULL;
182 
183 	if (dma_release_from_coherent(dev, order, vaddr))
184 		return;
185 
186 	plat_unmap_dma_mem(dev, dma_handle, size, DMA_BIDIRECTIONAL);
187 
188 	if (!plat_device_is_coherent(dev) && !hw_coherentio)
189 		addr = CAC_ADDR(addr);
190 
191 	page = virt_to_page((void *) addr);
192 
193 	if (!dma_release_from_contiguous(dev, page, count))
194 		__free_pages(page, get_order(size));
195 }
196 
__dma_sync_virtual(void * addr,size_t size,enum dma_data_direction direction)197 static inline void __dma_sync_virtual(void *addr, size_t size,
198 	enum dma_data_direction direction)
199 {
200 	switch (direction) {
201 	case DMA_TO_DEVICE:
202 		dma_cache_wback((unsigned long)addr, size);
203 		break;
204 
205 	case DMA_FROM_DEVICE:
206 		dma_cache_inv((unsigned long)addr, size);
207 		break;
208 
209 	case DMA_BIDIRECTIONAL:
210 		dma_cache_wback_inv((unsigned long)addr, size);
211 		break;
212 
213 	default:
214 		BUG();
215 	}
216 }
217 
218 /*
219  * A single sg entry may refer to multiple physically contiguous
220  * pages. But we still need to process highmem pages individually.
221  * If highmem is not configured then the bulk of this loop gets
222  * optimized out.
223  */
__dma_sync(struct page * page,unsigned long offset,size_t size,enum dma_data_direction direction)224 static inline void __dma_sync(struct page *page,
225 	unsigned long offset, size_t size, enum dma_data_direction direction)
226 {
227 	size_t left = size;
228 
229 	do {
230 		size_t len = left;
231 
232 		if (PageHighMem(page)) {
233 			void *addr;
234 
235 			if (offset + len > PAGE_SIZE) {
236 				if (offset >= PAGE_SIZE) {
237 					page += offset >> PAGE_SHIFT;
238 					offset &= ~PAGE_MASK;
239 				}
240 				len = PAGE_SIZE - offset;
241 			}
242 
243 			addr = kmap_atomic(page);
244 			__dma_sync_virtual(addr + offset, len, direction);
245 			kunmap_atomic(addr);
246 		} else
247 			__dma_sync_virtual(page_address(page) + offset,
248 					   size, direction);
249 		offset = 0;
250 		page++;
251 		left -= len;
252 	} while (left);
253 }
254 
mips_dma_unmap_page(struct device * dev,dma_addr_t dma_addr,size_t size,enum dma_data_direction direction,struct dma_attrs * attrs)255 static void mips_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
256 	size_t size, enum dma_data_direction direction, struct dma_attrs *attrs)
257 {
258 	if (cpu_needs_post_dma_flush(dev))
259 		__dma_sync(dma_addr_to_page(dev, dma_addr),
260 			   dma_addr & ~PAGE_MASK, size, direction);
261 	plat_post_dma_flush(dev);
262 	plat_unmap_dma_mem(dev, dma_addr, size, direction);
263 }
264 
mips_dma_map_sg(struct device * dev,struct scatterlist * sg,int nents,enum dma_data_direction direction,struct dma_attrs * attrs)265 static int mips_dma_map_sg(struct device *dev, struct scatterlist *sg,
266 	int nents, enum dma_data_direction direction, struct dma_attrs *attrs)
267 {
268 	int i;
269 
270 	for (i = 0; i < nents; i++, sg++) {
271 		if (!plat_device_is_coherent(dev))
272 			__dma_sync(sg_page(sg), sg->offset, sg->length,
273 				   direction);
274 #ifdef CONFIG_NEED_SG_DMA_LENGTH
275 		sg->dma_length = sg->length;
276 #endif
277 		sg->dma_address = plat_map_dma_mem_page(dev, sg_page(sg)) +
278 				  sg->offset;
279 	}
280 
281 	return nents;
282 }
283 
mips_dma_map_page(struct device * dev,struct page * page,unsigned long offset,size_t size,enum dma_data_direction direction,struct dma_attrs * attrs)284 static dma_addr_t mips_dma_map_page(struct device *dev, struct page *page,
285 	unsigned long offset, size_t size, enum dma_data_direction direction,
286 	struct dma_attrs *attrs)
287 {
288 	if (!plat_device_is_coherent(dev))
289 		__dma_sync(page, offset, size, direction);
290 
291 	return plat_map_dma_mem_page(dev, page) + offset;
292 }
293 
mips_dma_unmap_sg(struct device * dev,struct scatterlist * sg,int nhwentries,enum dma_data_direction direction,struct dma_attrs * attrs)294 static void mips_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
295 	int nhwentries, enum dma_data_direction direction,
296 	struct dma_attrs *attrs)
297 {
298 	int i;
299 
300 	for (i = 0; i < nhwentries; i++, sg++) {
301 		if (!plat_device_is_coherent(dev) &&
302 		    direction != DMA_TO_DEVICE)
303 			__dma_sync(sg_page(sg), sg->offset, sg->length,
304 				   direction);
305 		plat_unmap_dma_mem(dev, sg->dma_address, sg->length, direction);
306 	}
307 }
308 
mips_dma_sync_single_for_cpu(struct device * dev,dma_addr_t dma_handle,size_t size,enum dma_data_direction direction)309 static void mips_dma_sync_single_for_cpu(struct device *dev,
310 	dma_addr_t dma_handle, size_t size, enum dma_data_direction direction)
311 {
312 	if (cpu_needs_post_dma_flush(dev))
313 		__dma_sync(dma_addr_to_page(dev, dma_handle),
314 			   dma_handle & ~PAGE_MASK, size, direction);
315 	plat_post_dma_flush(dev);
316 }
317 
mips_dma_sync_single_for_device(struct device * dev,dma_addr_t dma_handle,size_t size,enum dma_data_direction direction)318 static void mips_dma_sync_single_for_device(struct device *dev,
319 	dma_addr_t dma_handle, size_t size, enum dma_data_direction direction)
320 {
321 	if (!plat_device_is_coherent(dev))
322 		__dma_sync(dma_addr_to_page(dev, dma_handle),
323 			   dma_handle & ~PAGE_MASK, size, direction);
324 }
325 
mips_dma_sync_sg_for_cpu(struct device * dev,struct scatterlist * sg,int nelems,enum dma_data_direction direction)326 static void mips_dma_sync_sg_for_cpu(struct device *dev,
327 	struct scatterlist *sg, int nelems, enum dma_data_direction direction)
328 {
329 	int i;
330 
331 	if (cpu_needs_post_dma_flush(dev))
332 		for (i = 0; i < nelems; i++, sg++)
333 			__dma_sync(sg_page(sg), sg->offset, sg->length,
334 				   direction);
335 	plat_post_dma_flush(dev);
336 }
337 
mips_dma_sync_sg_for_device(struct device * dev,struct scatterlist * sg,int nelems,enum dma_data_direction direction)338 static void mips_dma_sync_sg_for_device(struct device *dev,
339 	struct scatterlist *sg, int nelems, enum dma_data_direction direction)
340 {
341 	int i;
342 
343 	if (!plat_device_is_coherent(dev))
344 		for (i = 0; i < nelems; i++, sg++)
345 			__dma_sync(sg_page(sg), sg->offset, sg->length,
346 				   direction);
347 }
348 
mips_dma_mapping_error(struct device * dev,dma_addr_t dma_addr)349 int mips_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
350 {
351 	return 0;
352 }
353 
mips_dma_supported(struct device * dev,u64 mask)354 int mips_dma_supported(struct device *dev, u64 mask)
355 {
356 	return plat_dma_supported(dev, mask);
357 }
358 
dma_cache_sync(struct device * dev,void * vaddr,size_t size,enum dma_data_direction direction)359 void dma_cache_sync(struct device *dev, void *vaddr, size_t size,
360 			 enum dma_data_direction direction)
361 {
362 	BUG_ON(direction == DMA_NONE);
363 
364 	if (!plat_device_is_coherent(dev))
365 		__dma_sync_virtual(vaddr, size, direction);
366 }
367 
368 EXPORT_SYMBOL(dma_cache_sync);
369 
370 static struct dma_map_ops mips_default_dma_map_ops = {
371 	.alloc = mips_dma_alloc_coherent,
372 	.free = mips_dma_free_coherent,
373 	.map_page = mips_dma_map_page,
374 	.unmap_page = mips_dma_unmap_page,
375 	.map_sg = mips_dma_map_sg,
376 	.unmap_sg = mips_dma_unmap_sg,
377 	.sync_single_for_cpu = mips_dma_sync_single_for_cpu,
378 	.sync_single_for_device = mips_dma_sync_single_for_device,
379 	.sync_sg_for_cpu = mips_dma_sync_sg_for_cpu,
380 	.sync_sg_for_device = mips_dma_sync_sg_for_device,
381 	.mapping_error = mips_dma_mapping_error,
382 	.dma_supported = mips_dma_supported
383 };
384 
385 struct dma_map_ops *mips_dma_map_ops = &mips_default_dma_map_ops;
386 EXPORT_SYMBOL(mips_dma_map_ops);
387 
388 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
389 
mips_dma_init(void)390 static int __init mips_dma_init(void)
391 {
392 	dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
393 
394 	return 0;
395 }
396 fs_initcall(mips_dma_init);
397