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