1 // SPDX-License-Identifier: GPL-2.0 1 // SPDX-License-Identifier: GPL-2.0
2 /* 2 /*
3 * Copyright (C) 2018-2020 Christoph Hellwig. !! 3 * Copyright (C) 2018 Christoph Hellwig.
4 * 4 *
5 * DMA operations that map physical memory dir 5 * DMA operations that map physical memory directly without using an IOMMU.
6 */ 6 */
7 #include <linux/memblock.h> /* for max_pfn */ 7 #include <linux/memblock.h> /* for max_pfn */
8 #include <linux/export.h> 8 #include <linux/export.h>
9 #include <linux/mm.h> 9 #include <linux/mm.h>
10 #include <linux/dma-map-ops.h> !! 10 #include <linux/dma-direct.h>
11 #include <linux/scatterlist.h> 11 #include <linux/scatterlist.h>
>> 12 #include <linux/dma-contiguous.h>
>> 13 #include <linux/dma-noncoherent.h>
12 #include <linux/pfn.h> 14 #include <linux/pfn.h>
13 #include <linux/vmalloc.h> <<
14 #include <linux/set_memory.h> 15 #include <linux/set_memory.h>
15 #include <linux/slab.h> !! 16 #include <linux/swiotlb.h>
16 #include "direct.h" <<
17 17
18 /* 18 /*
19 * Most architectures use ZONE_DMA for the fir !! 19 * Most architectures use ZONE_DMA for the first 16 Megabytes, but
20 * it for entirely different regions. In that !! 20 * some use it for entirely different regions:
21 * override the variable below for dma-direct <<
22 */ 21 */
23 unsigned int zone_dma_bits __ro_after_init = 2 !! 22 #ifndef ARCH_ZONE_DMA_BITS
>> 23 #define ARCH_ZONE_DMA_BITS 24
>> 24 #endif
24 25
25 static inline dma_addr_t phys_to_dma_direct(st !! 26 /*
26 phys_addr_t phys) !! 27 * For AMD SEV all DMA must be to unencrypted addresses.
>> 28 */
>> 29 static inline bool force_dma_unencrypted(void)
27 { 30 {
28 if (force_dma_unencrypted(dev)) !! 31 return sev_active();
29 return phys_to_dma_unencrypted !! 32 }
30 return phys_to_dma(dev, phys); !! 33
>> 34 static void report_addr(struct device *dev, dma_addr_t dma_addr, size_t size)
>> 35 {
>> 36 if (!dev->dma_mask) {
>> 37 dev_err_once(dev, "DMA map on device without dma_mask\n");
>> 38 } else if (*dev->dma_mask >= DMA_BIT_MASK(32) || dev->bus_dma_mask) {
>> 39 dev_err_once(dev,
>> 40 "overflow %pad+%zu of DMA mask %llx bus mask %llx\n",
>> 41 &dma_addr, size, *dev->dma_mask, dev->bus_dma_mask);
>> 42 }
>> 43 WARN_ON_ONCE(1);
31 } 44 }
32 45
33 static inline struct page *dma_direct_to_page( !! 46 static inline dma_addr_t phys_to_dma_direct(struct device *dev,
34 dma_addr_t dma_addr) !! 47 phys_addr_t phys)
35 { 48 {
36 return pfn_to_page(PHYS_PFN(dma_to_phy !! 49 if (force_dma_unencrypted())
>> 50 return __phys_to_dma(dev, phys);
>> 51 return phys_to_dma(dev, phys);
37 } 52 }
38 53
39 u64 dma_direct_get_required_mask(struct device 54 u64 dma_direct_get_required_mask(struct device *dev)
40 { 55 {
41 phys_addr_t phys = (phys_addr_t)(max_p !! 56 u64 max_dma = phys_to_dma_direct(dev, (max_pfn - 1) << PAGE_SHIFT);
42 u64 max_dma = phys_to_dma_direct(dev, !! 57
>> 58 if (dev->bus_dma_mask && dev->bus_dma_mask < max_dma)
>> 59 max_dma = dev->bus_dma_mask;
43 60
44 return (1ULL << (fls64(max_dma) - 1)) 61 return (1ULL << (fls64(max_dma) - 1)) * 2 - 1;
45 } 62 }
46 63
47 static gfp_t dma_direct_optimal_gfp_mask(struc !! 64 static gfp_t __dma_direct_optimal_gfp_mask(struct device *dev, u64 dma_mask,
>> 65 u64 *phys_mask)
48 { 66 {
49 u64 dma_limit = min_not_zero( !! 67 if (dev->bus_dma_mask && dev->bus_dma_mask < dma_mask)
50 dev->coherent_dma_mask, !! 68 dma_mask = dev->bus_dma_mask;
51 dev->bus_dma_limit); !! 69
>> 70 if (force_dma_unencrypted())
>> 71 *phys_mask = __dma_to_phys(dev, dma_mask);
>> 72 else
>> 73 *phys_mask = dma_to_phys(dev, dma_mask);
52 74
53 /* 75 /*
54 * Optimistically try the zone that th 76 * Optimistically try the zone that the physical address mask falls
55 * into first. If that returns memory 77 * into first. If that returns memory that isn't actually addressable
56 * we will fallback to the next lower 78 * we will fallback to the next lower zone and try again.
57 * 79 *
58 * Note that GFP_DMA32 and GFP_DMA are 80 * Note that GFP_DMA32 and GFP_DMA are no ops without the corresponding
59 * zones. 81 * zones.
60 */ 82 */
61 *phys_limit = dma_to_phys(dev, dma_lim !! 83 if (*phys_mask <= DMA_BIT_MASK(ARCH_ZONE_DMA_BITS))
62 if (*phys_limit <= DMA_BIT_MASK(zone_d <<
63 return GFP_DMA; 84 return GFP_DMA;
64 if (*phys_limit <= DMA_BIT_MASK(32)) !! 85 if (*phys_mask <= DMA_BIT_MASK(32))
65 return GFP_DMA32; 86 return GFP_DMA32;
66 return 0; 87 return 0;
67 } 88 }
68 89
69 bool dma_coherent_ok(struct device *dev, phys_ !! 90 static bool dma_coherent_ok(struct device *dev, phys_addr_t phys, size_t size)
70 { <<
71 dma_addr_t dma_addr = phys_to_dma_dire <<
72 <<
73 if (dma_addr == DMA_MAPPING_ERROR) <<
74 return false; <<
75 return dma_addr + size - 1 <= <<
76 min_not_zero(dev->coherent_dma <<
77 } <<
78 <<
79 static int dma_set_decrypted(struct device *de <<
80 { <<
81 if (!force_dma_unencrypted(dev)) <<
82 return 0; <<
83 return set_memory_decrypted((unsigned <<
84 } <<
85 <<
86 static int dma_set_encrypted(struct device *de <<
87 { <<
88 int ret; <<
89 <<
90 if (!force_dma_unencrypted(dev)) <<
91 return 0; <<
92 ret = set_memory_encrypted((unsigned l <<
93 if (ret) <<
94 pr_warn_ratelimited("leaking D <<
95 return ret; <<
96 } <<
97 <<
98 static void __dma_direct_free_pages(struct dev <<
99 size_t siz <<
100 { 91 {
101 if (swiotlb_free(dev, page, size)) !! 92 return phys_to_dma_direct(dev, phys) + size - 1 <=
102 return; !! 93 min_not_zero(dev->coherent_dma_mask, dev->bus_dma_mask);
103 dma_free_contiguous(dev, page, size); <<
104 } 94 }
105 95
106 static struct page *dma_direct_alloc_swiotlb(s !! 96 struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
107 { !! 97 dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
108 struct page *page = swiotlb_alloc(dev, <<
109 <<
110 if (page && !dma_coherent_ok(dev, page <<
111 swiotlb_free(dev, page, size); <<
112 return NULL; <<
113 } <<
114 <<
115 return page; <<
116 } <<
117 <<
118 static struct page *__dma_direct_alloc_pages(s <<
119 gfp_t gfp, bool allow_highmem) <<
120 { 98 {
121 int node = dev_to_node(dev); !! 99 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
>> 100 int page_order = get_order(size);
122 struct page *page = NULL; 101 struct page *page = NULL;
123 u64 phys_limit; !! 102 u64 phys_mask;
124 <<
125 WARN_ON_ONCE(!PAGE_ALIGNED(size)); <<
126 103
127 if (is_swiotlb_for_alloc(dev)) !! 104 if (attrs & DMA_ATTR_NO_WARN)
128 return dma_direct_alloc_swiotl !! 105 gfp |= __GFP_NOWARN;
129 106
130 gfp |= dma_direct_optimal_gfp_mask(dev !! 107 /* we always manually zero the memory once we are done: */
131 page = dma_alloc_contiguous(dev, size, !! 108 gfp &= ~__GFP_ZERO;
132 if (page) { !! 109 gfp |= __dma_direct_optimal_gfp_mask(dev, dev->coherent_dma_mask,
133 if (!dma_coherent_ok(dev, page !! 110 &phys_mask);
134 (!allow_highmem && PageHig !! 111 again:
135 dma_free_contiguous(de !! 112 /* CMA can be used only in the context which permits sleeping */
>> 113 if (gfpflags_allow_blocking(gfp)) {
>> 114 page = dma_alloc_from_contiguous(dev, count, page_order,
>> 115 gfp & __GFP_NOWARN);
>> 116 if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
>> 117 dma_release_from_contiguous(dev, page, count);
136 page = NULL; 118 page = NULL;
137 } 119 }
138 } 120 }
139 again: <<
140 if (!page) 121 if (!page)
141 page = alloc_pages_node(node, !! 122 page = alloc_pages_node(dev_to_node(dev), gfp, page_order);
>> 123
142 if (page && !dma_coherent_ok(dev, page 124 if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
143 dma_free_contiguous(dev, page, !! 125 __free_pages(page, page_order);
144 page = NULL; 126 page = NULL;
145 127
146 if (IS_ENABLED(CONFIG_ZONE_DMA 128 if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
147 phys_limit < DMA_BIT_MASK( !! 129 phys_mask < DMA_BIT_MASK(64) &&
148 !(gfp & (GFP_DMA32 | GFP_D 130 !(gfp & (GFP_DMA32 | GFP_DMA))) {
149 gfp |= GFP_DMA32; 131 gfp |= GFP_DMA32;
150 goto again; 132 goto again;
151 } 133 }
152 134
153 if (IS_ENABLED(CONFIG_ZONE_DMA 135 if (IS_ENABLED(CONFIG_ZONE_DMA) && !(gfp & GFP_DMA)) {
154 gfp = (gfp & ~GFP_DMA3 136 gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
155 goto again; 137 goto again;
156 } 138 }
157 } 139 }
158 140
159 return page; 141 return page;
160 } 142 }
161 143
162 /* !! 144 void *dma_direct_alloc_pages(struct device *dev, size_t size,
163 * Check if a potentially blocking operations !! 145 dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
164 * pools for the given device/gfp. <<
165 */ <<
166 static bool dma_direct_use_pool(struct device <<
167 { <<
168 return !gfpflags_allow_blocking(gfp) & <<
169 } <<
170 <<
171 static void *dma_direct_alloc_from_pool(struct <<
172 dma_addr_t *dma_handle, gfp_t <<
173 { 146 {
174 struct page *page; 147 struct page *page;
175 u64 phys_limit; <<
176 void *ret; 148 void *ret;
177 149
178 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_DM !! 150 page = __dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
179 return NULL; <<
180 <<
181 gfp |= dma_direct_optimal_gfp_mask(dev <<
182 page = dma_alloc_from_pool(dev, size, <<
183 if (!page) <<
184 return NULL; <<
185 *dma_handle = phys_to_dma_direct(dev, <<
186 return ret; <<
187 } <<
188 <<
189 static void *dma_direct_alloc_no_mapping(struc <<
190 dma_addr_t *dma_handle, gfp_t <<
191 { <<
192 struct page *page; <<
193 <<
194 page = __dma_direct_alloc_pages(dev, s <<
195 if (!page) 151 if (!page)
196 return NULL; 152 return NULL;
197 153
198 /* remove any dirty cache lines on the !! 154 if (PageHighMem(page)) {
199 if (!PageHighMem(page)) <<
200 arch_dma_prep_coherent(page, s <<
201 <<
202 /* return the page pointer as the opaq <<
203 *dma_handle = phys_to_dma_direct(dev, <<
204 return page; <<
205 } <<
206 <<
207 void *dma_direct_alloc(struct device *dev, siz <<
208 dma_addr_t *dma_handle, gfp_t <<
209 { <<
210 bool remap = false, set_uncached = fal <<
211 struct page *page; <<
212 void *ret; <<
213 <<
214 size = PAGE_ALIGN(size); <<
215 if (attrs & DMA_ATTR_NO_WARN) <<
216 gfp |= __GFP_NOWARN; <<
217 <<
218 if ((attrs & DMA_ATTR_NO_KERNEL_MAPPIN <<
219 !force_dma_unencrypted(dev) && !is <<
220 return dma_direct_alloc_no_map <<
221 <<
222 if (!dev_is_dma_coherent(dev)) { <<
223 if (IS_ENABLED(CONFIG_ARCH_HAS <<
224 !is_swiotlb_for_alloc(dev) <<
225 return arch_dma_alloc( <<
226 <<
227 <<
228 /* <<
229 * If there is a global pool, <<
230 * non-coherent devices. <<
231 */ <<
232 if (IS_ENABLED(CONFIG_DMA_GLOB <<
233 return dma_alloc_from_ <<
234 dma_ha <<
235 <<
236 /* 155 /*
237 * Otherwise we require the ar !! 156 * Depending on the cma= arguments and per-arch setup
238 * mark arbitrary parts of the !! 157 * dma_alloc_from_contiguous could return highmem pages.
239 * or remapped it uncached. !! 158 * Without remapping there is no way to return them here,
>> 159 * so log an error and fail.
240 */ 160 */
241 set_uncached = IS_ENABLED(CONF !! 161 dev_info(dev, "Rejecting highmem page from CMA.\n");
242 remap = IS_ENABLED(CONFIG_DMA_ !! 162 __dma_direct_free_pages(dev, size, page);
243 if (!set_uncached && !remap) { <<
244 pr_warn_once("coherent <<
245 return NULL; <<
246 } <<
247 } <<
248 <<
249 /* <<
250 * Remapping or decrypting memory may <<
251 * the atomic pools instead if we aren <<
252 */ <<
253 if ((remap || force_dma_unencrypted(de <<
254 dma_direct_use_pool(dev, gfp)) <<
255 return dma_direct_alloc_from_p <<
256 <<
257 /* we always manually zero the memory <<
258 page = __dma_direct_alloc_pages(dev, s <<
259 if (!page) <<
260 return NULL; 163 return NULL;
261 <<
262 /* <<
263 * dma_alloc_contiguous can return hig <<
264 * combination the cma= arguments and <<
265 * remapped to return a kernel virtual <<
266 */ <<
267 if (PageHighMem(page)) { <<
268 remap = true; <<
269 set_uncached = false; <<
270 } 164 }
271 165
272 if (remap) { !! 166 ret = page_address(page);
273 pgprot_t prot = dma_pgprot(dev !! 167 if (force_dma_unencrypted()) {
274 !! 168 set_memory_decrypted((unsigned long)ret, 1 << get_order(size));
275 if (force_dma_unencrypted(dev) !! 169 *dma_handle = __phys_to_dma(dev, page_to_phys(page));
276 prot = pgprot_decrypte <<
277 <<
278 /* remove any dirty cache line <<
279 arch_dma_prep_coherent(page, s <<
280 <<
281 /* create a coherent mapping * <<
282 ret = dma_common_contiguous_re <<
283 __builtin_retu <<
284 if (!ret) <<
285 goto out_free_pages; <<
286 } else { 170 } else {
287 ret = page_address(page); !! 171 *dma_handle = phys_to_dma(dev, page_to_phys(page));
288 if (dma_set_decrypted(dev, ret <<
289 goto out_leak_pages; <<
290 } 172 }
291 <<
292 memset(ret, 0, size); 173 memset(ret, 0, size);
293 <<
294 if (set_uncached) { <<
295 arch_dma_prep_coherent(page, s <<
296 ret = arch_dma_set_uncached(re <<
297 if (IS_ERR(ret)) <<
298 goto out_encrypt_pages <<
299 } <<
300 <<
301 *dma_handle = phys_to_dma_direct(dev, <<
302 return ret; 174 return ret;
303 <<
304 out_encrypt_pages: <<
305 if (dma_set_encrypted(dev, page_addres <<
306 return NULL; <<
307 out_free_pages: <<
308 __dma_direct_free_pages(dev, page, siz <<
309 return NULL; <<
310 out_leak_pages: <<
311 return NULL; <<
312 } 175 }
313 176
314 void dma_direct_free(struct device *dev, size_ !! 177 void __dma_direct_free_pages(struct device *dev, size_t size, struct page *page)
315 void *cpu_addr, dma_addr_t dma <<
316 { 178 {
317 unsigned int page_order = get_order(si !! 179 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
318 <<
319 if ((attrs & DMA_ATTR_NO_KERNEL_MAPPIN <<
320 !force_dma_unencrypted(dev) && !is <<
321 /* cpu_addr is a struct page c <<
322 dma_free_contiguous(dev, cpu_a <<
323 return; <<
324 } <<
325 <<
326 if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_ALL <<
327 !dev_is_dma_coherent(dev) && <<
328 !is_swiotlb_for_alloc(dev)) { <<
329 arch_dma_free(dev, size, cpu_a <<
330 return; <<
331 } <<
332 <<
333 if (IS_ENABLED(CONFIG_DMA_GLOBAL_POOL) <<
334 !dev_is_dma_coherent(dev)) { <<
335 if (!dma_release_from_global_c <<
336 WARN_ON_ONCE(1); <<
337 return; <<
338 } <<
339 <<
340 /* If cpu_addr is not from an atomic p <<
341 if (IS_ENABLED(CONFIG_DMA_COHERENT_POO <<
342 dma_free_from_pool(dev, cpu_addr, <<
343 return; <<
344 180
345 if (is_vmalloc_addr(cpu_addr)) { !! 181 if (!dma_release_from_contiguous(dev, page, count))
346 vunmap(cpu_addr); !! 182 __free_pages(page, get_order(size));
347 } else { <<
348 if (IS_ENABLED(CONFIG_ARCH_HAS <<
349 arch_dma_clear_uncache <<
350 if (dma_set_encrypted(dev, cpu <<
351 return; <<
352 } <<
353 <<
354 __dma_direct_free_pages(dev, dma_direc <<
355 } 183 }
356 184
357 struct page *dma_direct_alloc_pages(struct dev !! 185 void dma_direct_free_pages(struct device *dev, size_t size, void *cpu_addr,
358 dma_addr_t *dma_handle, enum d !! 186 dma_addr_t dma_addr, unsigned long attrs)
359 { 187 {
360 struct page *page; !! 188 unsigned int page_order = get_order(size);
361 void *ret; <<
362 189
363 if (force_dma_unencrypted(dev) && dma_ !! 190 if (force_dma_unencrypted())
364 return dma_direct_alloc_from_p !! 191 set_memory_encrypted((unsigned long)cpu_addr, 1 << page_order);
>> 192 __dma_direct_free_pages(dev, size, virt_to_page(cpu_addr));
>> 193 }
365 194
366 page = __dma_direct_alloc_pages(dev, s !! 195 void *dma_direct_alloc(struct device *dev, size_t size,
367 if (!page) !! 196 dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
368 return NULL; !! 197 {
>> 198 if (!dev_is_dma_coherent(dev))
>> 199 return arch_dma_alloc(dev, size, dma_handle, gfp, attrs);
>> 200 return dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
>> 201 }
369 202
370 ret = page_address(page); !! 203 void dma_direct_free(struct device *dev, size_t size,
371 if (dma_set_decrypted(dev, ret, size)) !! 204 void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs)
372 goto out_leak_pages; !! 205 {
373 memset(ret, 0, size); !! 206 if (!dev_is_dma_coherent(dev))
374 *dma_handle = phys_to_dma_direct(dev, !! 207 arch_dma_free(dev, size, cpu_addr, dma_addr, attrs);
375 return page; !! 208 else
376 out_leak_pages: !! 209 dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
377 return NULL; <<
378 } 210 }
379 211
380 void dma_direct_free_pages(struct device *dev, !! 212 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_DEVICE) || \
381 struct page *page, dma_addr_t !! 213 defined(CONFIG_SWIOTLB)
382 enum dma_data_direction dir) !! 214 void dma_direct_sync_single_for_device(struct device *dev,
>> 215 dma_addr_t addr, size_t size, enum dma_data_direction dir)
383 { 216 {
384 void *vaddr = page_address(page); !! 217 phys_addr_t paddr = dma_to_phys(dev, addr);
385 218
386 /* If cpu_addr is not from an atomic p !! 219 if (unlikely(is_swiotlb_buffer(paddr)))
387 if (IS_ENABLED(CONFIG_DMA_COHERENT_POO !! 220 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_DEVICE);
388 dma_free_from_pool(dev, vaddr, siz <<
389 return; <<
390 221
391 if (dma_set_encrypted(dev, vaddr, size !! 222 if (!dev_is_dma_coherent(dev))
392 return; !! 223 arch_sync_dma_for_device(dev, paddr, size, dir);
393 __dma_direct_free_pages(dev, page, siz <<
394 } 224 }
>> 225 EXPORT_SYMBOL(dma_direct_sync_single_for_device);
395 226
396 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_DEVIC <<
397 defined(CONFIG_SWIOTLB) <<
398 void dma_direct_sync_sg_for_device(struct devi 227 void dma_direct_sync_sg_for_device(struct device *dev,
399 struct scatterlist *sgl, int n 228 struct scatterlist *sgl, int nents, enum dma_data_direction dir)
400 { 229 {
401 struct scatterlist *sg; 230 struct scatterlist *sg;
402 int i; 231 int i;
403 232
404 for_each_sg(sgl, sg, nents, i) { 233 for_each_sg(sgl, sg, nents, i) {
405 phys_addr_t paddr = dma_to_phy !! 234 if (unlikely(is_swiotlb_buffer(sg_phys(sg))))
406 !! 235 swiotlb_tbl_sync_single(dev, sg_phys(sg), sg->length,
407 swiotlb_sync_single_for_device !! 236 dir, SYNC_FOR_DEVICE);
408 237
409 if (!dev_is_dma_coherent(dev)) 238 if (!dev_is_dma_coherent(dev))
410 arch_sync_dma_for_devi !! 239 arch_sync_dma_for_device(dev, sg_phys(sg), sg->length,
411 dir); 240 dir);
412 } 241 }
413 } 242 }
>> 243 EXPORT_SYMBOL(dma_direct_sync_sg_for_device);
414 #endif 244 #endif
415 245
416 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU) 246 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU) || \
417 defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_A 247 defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_ALL) || \
418 defined(CONFIG_SWIOTLB) 248 defined(CONFIG_SWIOTLB)
>> 249 void dma_direct_sync_single_for_cpu(struct device *dev,
>> 250 dma_addr_t addr, size_t size, enum dma_data_direction dir)
>> 251 {
>> 252 phys_addr_t paddr = dma_to_phys(dev, addr);
>> 253
>> 254 if (!dev_is_dma_coherent(dev)) {
>> 255 arch_sync_dma_for_cpu(dev, paddr, size, dir);
>> 256 arch_sync_dma_for_cpu_all(dev);
>> 257 }
>> 258
>> 259 if (unlikely(is_swiotlb_buffer(paddr)))
>> 260 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_CPU);
>> 261 }
>> 262 EXPORT_SYMBOL(dma_direct_sync_single_for_cpu);
>> 263
419 void dma_direct_sync_sg_for_cpu(struct device 264 void dma_direct_sync_sg_for_cpu(struct device *dev,
420 struct scatterlist *sgl, int n 265 struct scatterlist *sgl, int nents, enum dma_data_direction dir)
421 { 266 {
422 struct scatterlist *sg; 267 struct scatterlist *sg;
423 int i; 268 int i;
424 269
425 for_each_sg(sgl, sg, nents, i) { 270 for_each_sg(sgl, sg, nents, i) {
426 phys_addr_t paddr = dma_to_phy <<
427 <<
428 if (!dev_is_dma_coherent(dev)) 271 if (!dev_is_dma_coherent(dev))
429 arch_sync_dma_for_cpu( !! 272 arch_sync_dma_for_cpu(dev, sg_phys(sg), sg->length, dir);
430 !! 273
431 swiotlb_sync_single_for_cpu(de !! 274 if (unlikely(is_swiotlb_buffer(sg_phys(sg))))
432 !! 275 swiotlb_tbl_sync_single(dev, sg_phys(sg), sg->length, dir,
433 if (dir == DMA_FROM_DEVICE) !! 276 SYNC_FOR_CPU);
434 arch_dma_mark_clean(pa <<
435 } 277 }
436 278
437 if (!dev_is_dma_coherent(dev)) 279 if (!dev_is_dma_coherent(dev))
438 arch_sync_dma_for_cpu_all(); !! 280 arch_sync_dma_for_cpu_all(dev);
439 } 281 }
>> 282 EXPORT_SYMBOL(dma_direct_sync_sg_for_cpu);
>> 283
>> 284 void dma_direct_unmap_page(struct device *dev, dma_addr_t addr,
>> 285 size_t size, enum dma_data_direction dir, unsigned long attrs)
>> 286 {
>> 287 phys_addr_t phys = dma_to_phys(dev, addr);
>> 288
>> 289 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
>> 290 dma_direct_sync_single_for_cpu(dev, addr, size, dir);
>> 291
>> 292 if (unlikely(is_swiotlb_buffer(phys)))
>> 293 swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
>> 294 }
>> 295 EXPORT_SYMBOL(dma_direct_unmap_page);
440 296
441 /* <<
442 * Unmaps segments, except for ones marked as <<
443 * require any further action as they contain <<
444 */ <<
445 void dma_direct_unmap_sg(struct device *dev, s 297 void dma_direct_unmap_sg(struct device *dev, struct scatterlist *sgl,
446 int nents, enum dma_data_direc 298 int nents, enum dma_data_direction dir, unsigned long attrs)
447 { 299 {
448 struct scatterlist *sg; 300 struct scatterlist *sg;
449 int i; 301 int i;
450 302
451 for_each_sg(sgl, sg, nents, i) { !! 303 for_each_sg(sgl, sg, nents, i)
452 if (sg_dma_is_bus_address(sg)) !! 304 dma_direct_unmap_page(dev, sg->dma_address, sg_dma_len(sg), dir,
453 sg_dma_unmark_bus_addr !! 305 attrs);
454 else <<
455 dma_direct_unmap_page( <<
456 <<
457 } <<
458 } 306 }
>> 307 EXPORT_SYMBOL(dma_direct_unmap_sg);
459 #endif 308 #endif
460 309
>> 310 static inline bool dma_direct_possible(struct device *dev, dma_addr_t dma_addr,
>> 311 size_t size)
>> 312 {
>> 313 return swiotlb_force != SWIOTLB_FORCE &&
>> 314 (!dev || dma_capable(dev, dma_addr, size));
>> 315 }
>> 316
>> 317 dma_addr_t dma_direct_map_page(struct device *dev, struct page *page,
>> 318 unsigned long offset, size_t size, enum dma_data_direction dir,
>> 319 unsigned long attrs)
>> 320 {
>> 321 phys_addr_t phys = page_to_phys(page) + offset;
>> 322 dma_addr_t dma_addr = phys_to_dma(dev, phys);
>> 323
>> 324 if (unlikely(!dma_direct_possible(dev, dma_addr, size)) &&
>> 325 !swiotlb_map(dev, &phys, &dma_addr, size, dir, attrs)) {
>> 326 report_addr(dev, dma_addr, size);
>> 327 return DMA_MAPPING_ERROR;
>> 328 }
>> 329
>> 330 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
>> 331 arch_sync_dma_for_device(dev, phys, size, dir);
>> 332 return dma_addr;
>> 333 }
>> 334 EXPORT_SYMBOL(dma_direct_map_page);
>> 335
461 int dma_direct_map_sg(struct device *dev, stru 336 int dma_direct_map_sg(struct device *dev, struct scatterlist *sgl, int nents,
462 enum dma_data_direction dir, u 337 enum dma_data_direction dir, unsigned long attrs)
463 { 338 {
464 struct pci_p2pdma_map_state p2pdma_sta !! 339 int i;
465 enum pci_p2pdma_map_type map; <<
466 struct scatterlist *sg; 340 struct scatterlist *sg;
467 int i, ret; <<
468 341
469 for_each_sg(sgl, sg, nents, i) { 342 for_each_sg(sgl, sg, nents, i) {
470 if (is_pci_p2pdma_page(sg_page <<
471 map = pci_p2pdma_map_s <<
472 switch (map) { <<
473 case PCI_P2PDMA_MAP_BU <<
474 continue; <<
475 case PCI_P2PDMA_MAP_TH <<
476 /* <<
477 * Any P2P map <<
478 * host bridge <<
479 * address and <<
480 * done with d <<
481 */ <<
482 break; <<
483 default: <<
484 ret = -EREMOTE <<
485 goto out_unmap <<
486 } <<
487 } <<
488 <<
489 sg->dma_address = dma_direct_m 343 sg->dma_address = dma_direct_map_page(dev, sg_page(sg),
490 sg->offset, sg 344 sg->offset, sg->length, dir, attrs);
491 if (sg->dma_address == DMA_MAP !! 345 if (sg->dma_address == DMA_MAPPING_ERROR)
492 ret = -EIO; <<
493 goto out_unmap; 346 goto out_unmap;
494 } <<
495 sg_dma_len(sg) = sg->length; 347 sg_dma_len(sg) = sg->length;
496 } 348 }
497 349
498 return nents; 350 return nents;
499 351
500 out_unmap: 352 out_unmap:
501 dma_direct_unmap_sg(dev, sgl, i, dir, 353 dma_direct_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
502 return ret; !! 354 return 0;
503 } 355 }
>> 356 EXPORT_SYMBOL(dma_direct_map_sg);
504 357
505 dma_addr_t dma_direct_map_resource(struct devi 358 dma_addr_t dma_direct_map_resource(struct device *dev, phys_addr_t paddr,
506 size_t size, enum dma_data_dir 359 size_t size, enum dma_data_direction dir, unsigned long attrs)
507 { 360 {
508 dma_addr_t dma_addr = paddr; 361 dma_addr_t dma_addr = paddr;
509 362
510 if (unlikely(!dma_capable(dev, dma_add !! 363 if (unlikely(!dma_direct_possible(dev, dma_addr, size))) {
511 dev_err_once(dev, !! 364 report_addr(dev, dma_addr, size);
512 "DMA addr %pad+%z <<
513 &dma_addr, size, <<
514 WARN_ON_ONCE(1); <<
515 return DMA_MAPPING_ERROR; 365 return DMA_MAPPING_ERROR;
516 } 366 }
517 367
518 return dma_addr; 368 return dma_addr;
519 } 369 }
>> 370 EXPORT_SYMBOL(dma_direct_map_resource);
520 371
521 int dma_direct_get_sgtable(struct device *dev, !! 372 /*
522 void *cpu_addr, dma_addr_t dma !! 373 * Because 32-bit DMA masks are so common we expect every architecture to be
523 unsigned long attrs) !! 374 * able to satisfy them - either by not supporting more physical memory, or by
524 { !! 375 * providing a ZONE_DMA32. If neither is the case, the architecture needs to
525 struct page *page = dma_direct_to_page !! 376 * use an IOMMU instead of the direct mapping.
526 int ret; !! 377 */
527 <<
528 ret = sg_alloc_table(sgt, 1, GFP_KERNE <<
529 if (!ret) <<
530 sg_set_page(sgt->sgl, page, PA <<
531 return ret; <<
532 } <<
533 <<
534 bool dma_direct_can_mmap(struct device *dev) <<
535 { <<
536 return dev_is_dma_coherent(dev) || <<
537 IS_ENABLED(CONFIG_DMA_NONCOHER <<
538 } <<
539 <<
540 int dma_direct_mmap(struct device *dev, struct <<
541 void *cpu_addr, dma_addr_t dma <<
542 unsigned long attrs) <<
543 { <<
544 unsigned long user_count = vma_pages(v <<
545 unsigned long count = PAGE_ALIGN(size) <<
546 unsigned long pfn = PHYS_PFN(dma_to_ph <<
547 int ret = -ENXIO; <<
548 <<
549 vma->vm_page_prot = dma_pgprot(dev, vm <<
550 if (force_dma_unencrypted(dev)) <<
551 vma->vm_page_prot = pgprot_dec <<
552 <<
553 if (dma_mmap_from_dev_coherent(dev, vm <<
554 return ret; <<
555 if (dma_mmap_from_global_coherent(vma, <<
556 return ret; <<
557 <<
558 if (vma->vm_pgoff >= count || user_cou <<
559 return -ENXIO; <<
560 return remap_pfn_range(vma, vma->vm_st <<
561 user_count << PAGE_SHI <<
562 } <<
563 <<
564 int dma_direct_supported(struct device *dev, u 378 int dma_direct_supported(struct device *dev, u64 mask)
565 { 379 {
566 u64 min_mask = (max_pfn - 1) << PAGE_S !! 380 u64 min_mask;
567 381
568 /* !! 382 if (IS_ENABLED(CONFIG_ZONE_DMA))
569 * Because 32-bit DMA masks are so com !! 383 min_mask = DMA_BIT_MASK(ARCH_ZONE_DMA_BITS);
570 * to be able to satisfy them - either !! 384 else
571 * memory, or by providing a ZONE_DMA3 !! 385 min_mask = DMA_BIT_MASK(32);
572 * architecture needs to use an IOMMU !! 386
573 */ !! 387 min_mask = min_t(u64, min_mask, (max_pfn - 1) << PAGE_SHIFT);
574 if (mask >= DMA_BIT_MASK(32)) <<
575 return 1; <<
576 388
577 /* 389 /*
578 * This check needs to be against the !! 390 * This check needs to be against the actual bit mask value, so
579 * phys_to_dma_unencrypted() here so t !! 391 * use __phys_to_dma() here so that the SME encryption mask isn't
580 * part of the check. 392 * part of the check.
581 */ 393 */
582 if (IS_ENABLED(CONFIG_ZONE_DMA)) !! 394 return mask >= __phys_to_dma(dev, min_mask);
583 min_mask = min_t(u64, min_mask <<
584 return mask >= phys_to_dma_unencrypted <<
585 } <<
586 <<
587 /* <<
588 * To check whether all ram resource ranges ar <<
589 * Returns 0 when further check is needed <<
590 * Returns 1 if there is some RAM range can't <<
591 */ <<
592 static int check_ram_in_range_map(unsigned lon <<
593 unsigned lon <<
594 { <<
595 unsigned long end_pfn = start_pfn + nr <<
596 const struct bus_dma_region *bdr = NUL <<
597 const struct bus_dma_region *m; <<
598 struct device *dev = data; <<
599 <<
600 while (start_pfn < end_pfn) { <<
601 for (m = dev->dma_range_map; P <<
602 unsigned long cpu_star <<
603 <<
604 if (start_pfn >= cpu_s <<
605 start_pfn - cpu_st <<
606 bdr = m; <<
607 break; <<
608 } <<
609 } <<
610 if (!bdr) <<
611 return 1; <<
612 <<
613 start_pfn = PFN_DOWN(bdr->cpu_ <<
614 } <<
615 <<
616 return 0; <<
617 } <<
618 <<
619 bool dma_direct_all_ram_mapped(struct device * <<
620 { <<
621 if (!dev->dma_range_map) <<
622 return true; <<
623 return !walk_system_ram_range(0, PFN_D <<
624 check_ra <<
625 } 395 }
626 396
627 size_t dma_direct_max_mapping_size(struct devi 397 size_t dma_direct_max_mapping_size(struct device *dev)
628 { 398 {
629 /* If SWIOTLB is active, use its maxim !! 399 size_t size = SIZE_MAX;
630 if (is_swiotlb_active(dev) && <<
631 (dma_addressing_limited(dev) || is <<
632 return swiotlb_max_mapping_siz <<
633 return SIZE_MAX; <<
634 } <<
635 400
636 bool dma_direct_need_sync(struct device *dev, !! 401 /* If SWIOTLB is active, use its maximum mapping size */
637 { !! 402 if (is_swiotlb_active())
638 return !dev_is_dma_coherent(dev) || !! 403 size = swiotlb_max_mapping_size(dev);
639 swiotlb_find_pool(dev, dma_to_p <<
640 } <<
641 <<
642 /** <<
643 * dma_direct_set_offset - Assign scalar offse <<
644 * @dev: device pointer; needed to "own <<
645 * @cpu_start: beginning of memory region cov <<
646 * @dma_start: beginning of DMA/PCI region co <<
647 * @size: size of the region. <<
648 * <<
649 * This is for the simple case of a uniform of <<
650 * be discovered by "dma-ranges". <<
651 * <<
652 * It returns -ENOMEM if out of memory, -EINVA <<
653 * already exists, 0 otherwise. <<
654 * <<
655 * Note: any call to this from a driver is a b <<
656 * to be described by the device tree or other <<
657 */ <<
658 int dma_direct_set_offset(struct device *dev, <<
659 dma_addr_t dma_start, <<
660 { <<
661 struct bus_dma_region *map; <<
662 u64 offset = (u64)cpu_start - (u64)dma <<
663 404
664 if (dev->dma_range_map) { !! 405 return size;
665 dev_err(dev, "attempt to add D <<
666 return -EINVAL; <<
667 } <<
668 <<
669 if (!offset) <<
670 return 0; <<
671 <<
672 map = kcalloc(2, sizeof(*map), GFP_KER <<
673 if (!map) <<
674 return -ENOMEM; <<
675 map[0].cpu_start = cpu_start; <<
676 map[0].dma_start = dma_start; <<
677 map[0].size = size; <<
678 dev->dma_range_map = map; <<
679 return 0; <<
680 } 406 }
681 407
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.