Merge tag 'gcc-plugins-v4.9-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git...
[cascardo/linux.git] / arch / mips / mm / dma-default.c
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/export.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 #if defined(CONFIG_DMA_MAYBE_COHERENT) && !defined(CONFIG_DMA_PERDEV_COHERENT)
28 /* User defined DMA coherency from command line. */
29 enum coherent_io_user_state coherentio = IO_COHERENCE_DEFAULT;
30 EXPORT_SYMBOL_GPL(coherentio);
31 int hw_coherentio = 0;  /* Actual hardware supported DMA coherency setting. */
32
33 static int __init setcoherentio(char *str)
34 {
35         coherentio = IO_COHERENCE_ENABLED;
36         pr_info("Hardware DMA cache coherency (command line)\n");
37         return 0;
38 }
39 early_param("coherentio", setcoherentio);
40
41 static int __init setnocoherentio(char *str)
42 {
43         coherentio = IO_COHERENCE_DISABLED;
44         pr_info("Software DMA cache coherency (command line)\n");
45         return 0;
46 }
47 early_param("nocoherentio", setnocoherentio);
48 #endif
49
50 static inline struct page *dma_addr_to_page(struct device *dev,
51         dma_addr_t dma_addr)
52 {
53         return pfn_to_page(
54                 plat_dma_addr_to_phys(dev, dma_addr) >> PAGE_SHIFT);
55 }
56
57 /*
58  * The affected CPUs below in 'cpu_needs_post_dma_flush()' can
59  * speculatively fill random cachelines with stale data at any time,
60  * requiring an extra flush post-DMA.
61  *
62  * Warning on the terminology - Linux calls an uncached area coherent;
63  * MIPS terminology calls memory areas with hardware maintained coherency
64  * coherent.
65  *
66  * Note that the R14000 and R16000 should also be checked for in this
67  * condition.  However this function is only called on non-I/O-coherent
68  * systems and only the R10000 and R12000 are used in such systems, the
69  * SGI IP28 Indigo² rsp. SGI IP32 aka O2.
70  */
71 static inline int cpu_needs_post_dma_flush(struct device *dev)
72 {
73         return !plat_device_is_coherent(dev) &&
74                (boot_cpu_type() == CPU_R10000 ||
75                 boot_cpu_type() == CPU_R12000 ||
76                 boot_cpu_type() == CPU_BMIPS5000);
77 }
78
79 static gfp_t massage_gfp_flags(const struct device *dev, gfp_t gfp)
80 {
81         gfp_t dma_flag;
82
83         /* ignore region specifiers */
84         gfp &= ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM);
85
86 #ifdef CONFIG_ISA
87         if (dev == NULL)
88                 dma_flag = __GFP_DMA;
89         else
90 #endif
91 #if defined(CONFIG_ZONE_DMA32) && defined(CONFIG_ZONE_DMA)
92              if (dev == NULL || dev->coherent_dma_mask < DMA_BIT_MASK(32))
93                         dma_flag = __GFP_DMA;
94         else if (dev->coherent_dma_mask < DMA_BIT_MASK(64))
95                         dma_flag = __GFP_DMA32;
96         else
97 #endif
98 #if defined(CONFIG_ZONE_DMA32) && !defined(CONFIG_ZONE_DMA)
99              if (dev == NULL || dev->coherent_dma_mask < DMA_BIT_MASK(64))
100                 dma_flag = __GFP_DMA32;
101         else
102 #endif
103 #if defined(CONFIG_ZONE_DMA) && !defined(CONFIG_ZONE_DMA32)
104              if (dev == NULL ||
105                  dev->coherent_dma_mask < DMA_BIT_MASK(sizeof(phys_addr_t) * 8))
106                 dma_flag = __GFP_DMA;
107         else
108 #endif
109                 dma_flag = 0;
110
111         /* Don't invoke OOM killer */
112         gfp |= __GFP_NORETRY;
113
114         return gfp | dma_flag;
115 }
116
117 static void *mips_dma_alloc_noncoherent(struct device *dev, size_t size,
118         dma_addr_t * dma_handle, gfp_t gfp)
119 {
120         void *ret;
121
122         gfp = massage_gfp_flags(dev, gfp);
123
124         ret = (void *) __get_free_pages(gfp, get_order(size));
125
126         if (ret != NULL) {
127                 memset(ret, 0, size);
128                 *dma_handle = plat_map_dma_mem(dev, ret, size);
129         }
130
131         return ret;
132 }
133
134 static void *mips_dma_alloc_coherent(struct device *dev, size_t size,
135         dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
136 {
137         void *ret;
138         struct page *page = NULL;
139         unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
140
141         /*
142          * XXX: seems like the coherent and non-coherent implementations could
143          * be consolidated.
144          */
145         if (attrs & DMA_ATTR_NON_CONSISTENT)
146                 return mips_dma_alloc_noncoherent(dev, size, dma_handle, gfp);
147
148         gfp = massage_gfp_flags(dev, gfp);
149
150         if (IS_ENABLED(CONFIG_DMA_CMA) && gfpflags_allow_blocking(gfp))
151                 page = dma_alloc_from_contiguous(dev,
152                                         count, get_order(size));
153         if (!page)
154                 page = alloc_pages(gfp, get_order(size));
155
156         if (!page)
157                 return NULL;
158
159         ret = page_address(page);
160         memset(ret, 0, size);
161         *dma_handle = plat_map_dma_mem(dev, ret, size);
162         if (!plat_device_is_coherent(dev)) {
163                 dma_cache_wback_inv((unsigned long) ret, size);
164                 ret = UNCAC_ADDR(ret);
165         }
166
167         return ret;
168 }
169
170
171 static void mips_dma_free_noncoherent(struct device *dev, size_t size,
172                 void *vaddr, dma_addr_t dma_handle)
173 {
174         plat_unmap_dma_mem(dev, dma_handle, size, DMA_BIDIRECTIONAL);
175         free_pages((unsigned long) vaddr, get_order(size));
176 }
177
178 static void mips_dma_free_coherent(struct device *dev, size_t size, void *vaddr,
179         dma_addr_t dma_handle, unsigned long attrs)
180 {
181         unsigned long addr = (unsigned long) vaddr;
182         unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
183         struct page *page = NULL;
184
185         if (attrs & DMA_ATTR_NON_CONSISTENT) {
186                 mips_dma_free_noncoherent(dev, size, vaddr, dma_handle);
187                 return;
188         }
189
190         plat_unmap_dma_mem(dev, dma_handle, size, DMA_BIDIRECTIONAL);
191
192         if (!plat_device_is_coherent(dev))
193                 addr = CAC_ADDR(addr);
194
195         page = virt_to_page((void *) addr);
196
197         if (!dma_release_from_contiguous(dev, page, count))
198                 __free_pages(page, get_order(size));
199 }
200
201 static int mips_dma_mmap(struct device *dev, struct vm_area_struct *vma,
202         void *cpu_addr, dma_addr_t dma_addr, size_t size,
203         unsigned long attrs)
204 {
205         unsigned long user_count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
206         unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT;
207         unsigned long addr = (unsigned long)cpu_addr;
208         unsigned long off = vma->vm_pgoff;
209         unsigned long pfn;
210         int ret = -ENXIO;
211
212         if (!plat_device_is_coherent(dev))
213                 addr = CAC_ADDR(addr);
214
215         pfn = page_to_pfn(virt_to_page((void *)addr));
216
217         if (attrs & DMA_ATTR_WRITE_COMBINE)
218                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
219         else
220                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
221
222         if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
223                 return ret;
224
225         if (off < count && user_count <= (count - off)) {
226                 ret = remap_pfn_range(vma, vma->vm_start,
227                                       pfn + off,
228                                       user_count << PAGE_SHIFT,
229                                       vma->vm_page_prot);
230         }
231
232         return ret;
233 }
234
235 static inline void __dma_sync_virtual(void *addr, size_t size,
236         enum dma_data_direction direction)
237 {
238         switch (direction) {
239         case DMA_TO_DEVICE:
240                 dma_cache_wback((unsigned long)addr, size);
241                 break;
242
243         case DMA_FROM_DEVICE:
244                 dma_cache_inv((unsigned long)addr, size);
245                 break;
246
247         case DMA_BIDIRECTIONAL:
248                 dma_cache_wback_inv((unsigned long)addr, size);
249                 break;
250
251         default:
252                 BUG();
253         }
254 }
255
256 /*
257  * A single sg entry may refer to multiple physically contiguous
258  * pages. But we still need to process highmem pages individually.
259  * If highmem is not configured then the bulk of this loop gets
260  * optimized out.
261  */
262 static inline void __dma_sync(struct page *page,
263         unsigned long offset, size_t size, enum dma_data_direction direction)
264 {
265         size_t left = size;
266
267         do {
268                 size_t len = left;
269
270                 if (PageHighMem(page)) {
271                         void *addr;
272
273                         if (offset + len > PAGE_SIZE) {
274                                 if (offset >= PAGE_SIZE) {
275                                         page += offset >> PAGE_SHIFT;
276                                         offset &= ~PAGE_MASK;
277                                 }
278                                 len = PAGE_SIZE - offset;
279                         }
280
281                         addr = kmap_atomic(page);
282                         __dma_sync_virtual(addr + offset, len, direction);
283                         kunmap_atomic(addr);
284                 } else
285                         __dma_sync_virtual(page_address(page) + offset,
286                                            size, direction);
287                 offset = 0;
288                 page++;
289                 left -= len;
290         } while (left);
291 }
292
293 static void mips_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
294         size_t size, enum dma_data_direction direction, unsigned long attrs)
295 {
296         if (cpu_needs_post_dma_flush(dev))
297                 __dma_sync(dma_addr_to_page(dev, dma_addr),
298                            dma_addr & ~PAGE_MASK, size, direction);
299         plat_post_dma_flush(dev);
300         plat_unmap_dma_mem(dev, dma_addr, size, direction);
301 }
302
303 static int mips_dma_map_sg(struct device *dev, struct scatterlist *sglist,
304         int nents, enum dma_data_direction direction, unsigned long attrs)
305 {
306         int i;
307         struct scatterlist *sg;
308
309         for_each_sg(sglist, sg, nents, i) {
310                 if (!plat_device_is_coherent(dev))
311                         __dma_sync(sg_page(sg), sg->offset, sg->length,
312                                    direction);
313 #ifdef CONFIG_NEED_SG_DMA_LENGTH
314                 sg->dma_length = sg->length;
315 #endif
316                 sg->dma_address = plat_map_dma_mem_page(dev, sg_page(sg)) +
317                                   sg->offset;
318         }
319
320         return nents;
321 }
322
323 static dma_addr_t mips_dma_map_page(struct device *dev, struct page *page,
324         unsigned long offset, size_t size, enum dma_data_direction direction,
325         unsigned long attrs)
326 {
327         if (!plat_device_is_coherent(dev))
328                 __dma_sync(page, offset, size, direction);
329
330         return plat_map_dma_mem_page(dev, page) + offset;
331 }
332
333 static void mips_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
334         int nhwentries, enum dma_data_direction direction,
335         unsigned long attrs)
336 {
337         int i;
338         struct scatterlist *sg;
339
340         for_each_sg(sglist, sg, nhwentries, i) {
341                 if (!plat_device_is_coherent(dev) &&
342                     direction != DMA_TO_DEVICE)
343                         __dma_sync(sg_page(sg), sg->offset, sg->length,
344                                    direction);
345                 plat_unmap_dma_mem(dev, sg->dma_address, sg->length, direction);
346         }
347 }
348
349 static void mips_dma_sync_single_for_cpu(struct device *dev,
350         dma_addr_t dma_handle, size_t size, enum dma_data_direction direction)
351 {
352         if (cpu_needs_post_dma_flush(dev))
353                 __dma_sync(dma_addr_to_page(dev, dma_handle),
354                            dma_handle & ~PAGE_MASK, size, direction);
355         plat_post_dma_flush(dev);
356 }
357
358 static void mips_dma_sync_single_for_device(struct device *dev,
359         dma_addr_t dma_handle, size_t size, enum dma_data_direction direction)
360 {
361         if (!plat_device_is_coherent(dev))
362                 __dma_sync(dma_addr_to_page(dev, dma_handle),
363                            dma_handle & ~PAGE_MASK, size, direction);
364 }
365
366 static void mips_dma_sync_sg_for_cpu(struct device *dev,
367         struct scatterlist *sglist, int nelems,
368         enum dma_data_direction direction)
369 {
370         int i;
371         struct scatterlist *sg;
372
373         if (cpu_needs_post_dma_flush(dev)) {
374                 for_each_sg(sglist, sg, nelems, i) {
375                         __dma_sync(sg_page(sg), sg->offset, sg->length,
376                                    direction);
377                 }
378         }
379         plat_post_dma_flush(dev);
380 }
381
382 static void mips_dma_sync_sg_for_device(struct device *dev,
383         struct scatterlist *sglist, int nelems,
384         enum dma_data_direction direction)
385 {
386         int i;
387         struct scatterlist *sg;
388
389         if (!plat_device_is_coherent(dev)) {
390                 for_each_sg(sglist, sg, nelems, i) {
391                         __dma_sync(sg_page(sg), sg->offset, sg->length,
392                                    direction);
393                 }
394         }
395 }
396
397 int mips_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
398 {
399         return 0;
400 }
401
402 int mips_dma_supported(struct device *dev, u64 mask)
403 {
404         return plat_dma_supported(dev, mask);
405 }
406
407 void dma_cache_sync(struct device *dev, void *vaddr, size_t size,
408                          enum dma_data_direction direction)
409 {
410         BUG_ON(direction == DMA_NONE);
411
412         if (!plat_device_is_coherent(dev))
413                 __dma_sync_virtual(vaddr, size, direction);
414 }
415
416 EXPORT_SYMBOL(dma_cache_sync);
417
418 static struct dma_map_ops mips_default_dma_map_ops = {
419         .alloc = mips_dma_alloc_coherent,
420         .free = mips_dma_free_coherent,
421         .mmap = mips_dma_mmap,
422         .map_page = mips_dma_map_page,
423         .unmap_page = mips_dma_unmap_page,
424         .map_sg = mips_dma_map_sg,
425         .unmap_sg = mips_dma_unmap_sg,
426         .sync_single_for_cpu = mips_dma_sync_single_for_cpu,
427         .sync_single_for_device = mips_dma_sync_single_for_device,
428         .sync_sg_for_cpu = mips_dma_sync_sg_for_cpu,
429         .sync_sg_for_device = mips_dma_sync_sg_for_device,
430         .mapping_error = mips_dma_mapping_error,
431         .dma_supported = mips_dma_supported
432 };
433
434 struct dma_map_ops *mips_dma_map_ops = &mips_default_dma_map_ops;
435 EXPORT_SYMBOL(mips_dma_map_ops);
436
437 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
438
439 static int __init mips_dma_init(void)
440 {
441         dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
442
443         return 0;
444 }
445 fs_initcall(mips_dma_init);