xref: /openbmc/linux/kernel/dma/direct.c (revision cb1aaebe)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2018 Christoph Hellwig.
4  *
5  * DMA operations that map physical memory directly without using an IOMMU.
6  */
7 #include <linux/memblock.h> /* for max_pfn */
8 #include <linux/export.h>
9 #include <linux/mm.h>
10 #include <linux/dma-direct.h>
11 #include <linux/scatterlist.h>
12 #include <linux/dma-contiguous.h>
13 #include <linux/dma-noncoherent.h>
14 #include <linux/pfn.h>
15 #include <linux/set_memory.h>
16 #include <linux/swiotlb.h>
17 
18 /*
19  * Most architectures use ZONE_DMA for the first 16 Megabytes, but
20  * some use it for entirely different regions:
21  */
22 #ifndef ARCH_ZONE_DMA_BITS
23 #define ARCH_ZONE_DMA_BITS 24
24 #endif
25 
26 /*
27  * For AMD SEV all DMA must be to unencrypted addresses.
28  */
29 static inline bool force_dma_unencrypted(void)
30 {
31 	return sev_active();
32 }
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);
44 }
45 
46 static inline dma_addr_t phys_to_dma_direct(struct device *dev,
47 		phys_addr_t phys)
48 {
49 	if (force_dma_unencrypted())
50 		return __phys_to_dma(dev, phys);
51 	return phys_to_dma(dev, phys);
52 }
53 
54 u64 dma_direct_get_required_mask(struct device *dev)
55 {
56 	u64 max_dma = phys_to_dma_direct(dev, (max_pfn - 1) << PAGE_SHIFT);
57 
58 	if (dev->bus_dma_mask && dev->bus_dma_mask < max_dma)
59 		max_dma = dev->bus_dma_mask;
60 
61 	return (1ULL << (fls64(max_dma) - 1)) * 2 - 1;
62 }
63 
64 static gfp_t __dma_direct_optimal_gfp_mask(struct device *dev, u64 dma_mask,
65 		u64 *phys_mask)
66 {
67 	if (dev->bus_dma_mask && dev->bus_dma_mask < dma_mask)
68 		dma_mask = dev->bus_dma_mask;
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);
74 
75 	/*
76 	 * Optimistically try the zone that the physical address mask falls
77 	 * into first.  If that returns memory that isn't actually addressable
78 	 * we will fallback to the next lower zone and try again.
79 	 *
80 	 * Note that GFP_DMA32 and GFP_DMA are no ops without the corresponding
81 	 * zones.
82 	 */
83 	if (*phys_mask <= DMA_BIT_MASK(ARCH_ZONE_DMA_BITS))
84 		return GFP_DMA;
85 	if (*phys_mask <= DMA_BIT_MASK(32))
86 		return GFP_DMA32;
87 	return 0;
88 }
89 
90 static bool dma_coherent_ok(struct device *dev, phys_addr_t phys, size_t size)
91 {
92 	return phys_to_dma_direct(dev, phys) + size - 1 <=
93 			min_not_zero(dev->coherent_dma_mask, dev->bus_dma_mask);
94 }
95 
96 struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
97 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
98 {
99 	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
100 	int page_order = get_order(size);
101 	struct page *page = NULL;
102 	u64 phys_mask;
103 
104 	if (attrs & DMA_ATTR_NO_WARN)
105 		gfp |= __GFP_NOWARN;
106 
107 	/* we always manually zero the memory once we are done: */
108 	gfp &= ~__GFP_ZERO;
109 	gfp |= __dma_direct_optimal_gfp_mask(dev, dev->coherent_dma_mask,
110 			&phys_mask);
111 again:
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);
118 			page = NULL;
119 		}
120 	}
121 	if (!page)
122 		page = alloc_pages_node(dev_to_node(dev), gfp, page_order);
123 
124 	if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
125 		__free_pages(page, page_order);
126 		page = NULL;
127 
128 		if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
129 		    phys_mask < DMA_BIT_MASK(64) &&
130 		    !(gfp & (GFP_DMA32 | GFP_DMA))) {
131 			gfp |= GFP_DMA32;
132 			goto again;
133 		}
134 
135 		if (IS_ENABLED(CONFIG_ZONE_DMA) && !(gfp & GFP_DMA)) {
136 			gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
137 			goto again;
138 		}
139 	}
140 
141 	return page;
142 }
143 
144 void *dma_direct_alloc_pages(struct device *dev, size_t size,
145 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
146 {
147 	struct page *page;
148 	void *ret;
149 
150 	page = __dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
151 	if (!page)
152 		return NULL;
153 
154 	if (PageHighMem(page)) {
155 		/*
156 		 * Depending on the cma= arguments and per-arch setup
157 		 * dma_alloc_from_contiguous could return highmem pages.
158 		 * Without remapping there is no way to return them here,
159 		 * so log an error and fail.
160 		 */
161 		dev_info(dev, "Rejecting highmem page from CMA.\n");
162 		__dma_direct_free_pages(dev, size, page);
163 		return NULL;
164 	}
165 
166 	ret = page_address(page);
167 	if (force_dma_unencrypted()) {
168 		set_memory_decrypted((unsigned long)ret, 1 << get_order(size));
169 		*dma_handle = __phys_to_dma(dev, page_to_phys(page));
170 	} else {
171 		*dma_handle = phys_to_dma(dev, page_to_phys(page));
172 	}
173 	memset(ret, 0, size);
174 	return ret;
175 }
176 
177 void __dma_direct_free_pages(struct device *dev, size_t size, struct page *page)
178 {
179 	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
180 
181 	if (!dma_release_from_contiguous(dev, page, count))
182 		__free_pages(page, get_order(size));
183 }
184 
185 void dma_direct_free_pages(struct device *dev, size_t size, void *cpu_addr,
186 		dma_addr_t dma_addr, unsigned long attrs)
187 {
188 	unsigned int page_order = get_order(size);
189 
190 	if (force_dma_unencrypted())
191 		set_memory_encrypted((unsigned long)cpu_addr, 1 << page_order);
192 	__dma_direct_free_pages(dev, size, virt_to_page(cpu_addr));
193 }
194 
195 void *dma_direct_alloc(struct device *dev, size_t size,
196 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
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 }
202 
203 void dma_direct_free(struct device *dev, size_t size,
204 		void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs)
205 {
206 	if (!dev_is_dma_coherent(dev))
207 		arch_dma_free(dev, size, cpu_addr, dma_addr, attrs);
208 	else
209 		dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
210 }
211 
212 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_DEVICE) || \
213     defined(CONFIG_SWIOTLB)
214 void dma_direct_sync_single_for_device(struct device *dev,
215 		dma_addr_t addr, size_t size, enum dma_data_direction dir)
216 {
217 	phys_addr_t paddr = dma_to_phys(dev, addr);
218 
219 	if (unlikely(is_swiotlb_buffer(paddr)))
220 		swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_DEVICE);
221 
222 	if (!dev_is_dma_coherent(dev))
223 		arch_sync_dma_for_device(dev, paddr, size, dir);
224 }
225 EXPORT_SYMBOL(dma_direct_sync_single_for_device);
226 
227 void dma_direct_sync_sg_for_device(struct device *dev,
228 		struct scatterlist *sgl, int nents, enum dma_data_direction dir)
229 {
230 	struct scatterlist *sg;
231 	int i;
232 
233 	for_each_sg(sgl, sg, nents, i) {
234 		if (unlikely(is_swiotlb_buffer(sg_phys(sg))))
235 			swiotlb_tbl_sync_single(dev, sg_phys(sg), sg->length,
236 					dir, SYNC_FOR_DEVICE);
237 
238 		if (!dev_is_dma_coherent(dev))
239 			arch_sync_dma_for_device(dev, sg_phys(sg), sg->length,
240 					dir);
241 	}
242 }
243 EXPORT_SYMBOL(dma_direct_sync_sg_for_device);
244 #endif
245 
246 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU) || \
247     defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_ALL) || \
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 
264 void dma_direct_sync_sg_for_cpu(struct device *dev,
265 		struct scatterlist *sgl, int nents, enum dma_data_direction dir)
266 {
267 	struct scatterlist *sg;
268 	int i;
269 
270 	for_each_sg(sgl, sg, nents, i) {
271 		if (!dev_is_dma_coherent(dev))
272 			arch_sync_dma_for_cpu(dev, sg_phys(sg), sg->length, dir);
273 
274 		if (unlikely(is_swiotlb_buffer(sg_phys(sg))))
275 			swiotlb_tbl_sync_single(dev, sg_phys(sg), sg->length, dir,
276 					SYNC_FOR_CPU);
277 	}
278 
279 	if (!dev_is_dma_coherent(dev))
280 		arch_sync_dma_for_cpu_all(dev);
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);
296 
297 void dma_direct_unmap_sg(struct device *dev, struct scatterlist *sgl,
298 		int nents, enum dma_data_direction dir, unsigned long attrs)
299 {
300 	struct scatterlist *sg;
301 	int i;
302 
303 	for_each_sg(sgl, sg, nents, i)
304 		dma_direct_unmap_page(dev, sg->dma_address, sg_dma_len(sg), dir,
305 			     attrs);
306 }
307 EXPORT_SYMBOL(dma_direct_unmap_sg);
308 #endif
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 		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 
336 int dma_direct_map_sg(struct device *dev, struct scatterlist *sgl, int nents,
337 		enum dma_data_direction dir, unsigned long attrs)
338 {
339 	int i;
340 	struct scatterlist *sg;
341 
342 	for_each_sg(sgl, sg, nents, i) {
343 		sg->dma_address = dma_direct_map_page(dev, sg_page(sg),
344 				sg->offset, sg->length, dir, attrs);
345 		if (sg->dma_address == DMA_MAPPING_ERROR)
346 			goto out_unmap;
347 		sg_dma_len(sg) = sg->length;
348 	}
349 
350 	return nents;
351 
352 out_unmap:
353 	dma_direct_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
354 	return 0;
355 }
356 EXPORT_SYMBOL(dma_direct_map_sg);
357 
358 dma_addr_t dma_direct_map_resource(struct device *dev, phys_addr_t paddr,
359 		size_t size, enum dma_data_direction dir, unsigned long attrs)
360 {
361 	dma_addr_t dma_addr = paddr;
362 
363 	if (unlikely(!dma_direct_possible(dev, dma_addr, size))) {
364 		report_addr(dev, dma_addr, size);
365 		return DMA_MAPPING_ERROR;
366 	}
367 
368 	return dma_addr;
369 }
370 EXPORT_SYMBOL(dma_direct_map_resource);
371 
372 /*
373  * Because 32-bit DMA masks are so common we expect every architecture to be
374  * able to satisfy them - either by not supporting more physical memory, or by
375  * providing a ZONE_DMA32.  If neither is the case, the architecture needs to
376  * use an IOMMU instead of the direct mapping.
377  */
378 int dma_direct_supported(struct device *dev, u64 mask)
379 {
380 	u64 min_mask;
381 
382 	if (IS_ENABLED(CONFIG_ZONE_DMA))
383 		min_mask = DMA_BIT_MASK(ARCH_ZONE_DMA_BITS);
384 	else
385 		min_mask = DMA_BIT_MASK(32);
386 
387 	min_mask = min_t(u64, min_mask, (max_pfn - 1) << PAGE_SHIFT);
388 
389 	/*
390 	 * This check needs to be against the actual bit mask value, so
391 	 * use __phys_to_dma() here so that the SME encryption mask isn't
392 	 * part of the check.
393 	 */
394 	return mask >= __phys_to_dma(dev, min_mask);
395 }
396 
397 size_t dma_direct_max_mapping_size(struct device *dev)
398 {
399 	size_t size = SIZE_MAX;
400 
401 	/* If SWIOTLB is active, use its maximum mapping size */
402 	if (is_swiotlb_active())
403 		size = swiotlb_max_mapping_size(dev);
404 
405 	return size;
406 }
407