xref: /openbmc/linux/kernel/dma/direct.c (revision 2169e6da)
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 static void report_addr(struct device *dev, dma_addr_t dma_addr, size_t size)
27 {
28 	if (!dev->dma_mask) {
29 		dev_err_once(dev, "DMA map on device without dma_mask\n");
30 	} else if (*dev->dma_mask >= DMA_BIT_MASK(32) || dev->bus_dma_mask) {
31 		dev_err_once(dev,
32 			"overflow %pad+%zu of DMA mask %llx bus mask %llx\n",
33 			&dma_addr, size, *dev->dma_mask, dev->bus_dma_mask);
34 	}
35 	WARN_ON_ONCE(1);
36 }
37 
38 static inline dma_addr_t phys_to_dma_direct(struct device *dev,
39 		phys_addr_t phys)
40 {
41 	if (force_dma_unencrypted(dev))
42 		return __phys_to_dma(dev, phys);
43 	return phys_to_dma(dev, phys);
44 }
45 
46 u64 dma_direct_get_required_mask(struct device *dev)
47 {
48 	u64 max_dma = phys_to_dma_direct(dev, (max_pfn - 1) << PAGE_SHIFT);
49 
50 	if (dev->bus_dma_mask && dev->bus_dma_mask < max_dma)
51 		max_dma = dev->bus_dma_mask;
52 
53 	return (1ULL << (fls64(max_dma) - 1)) * 2 - 1;
54 }
55 
56 static gfp_t __dma_direct_optimal_gfp_mask(struct device *dev, u64 dma_mask,
57 		u64 *phys_mask)
58 {
59 	if (dev->bus_dma_mask && dev->bus_dma_mask < dma_mask)
60 		dma_mask = dev->bus_dma_mask;
61 
62 	if (force_dma_unencrypted(dev))
63 		*phys_mask = __dma_to_phys(dev, dma_mask);
64 	else
65 		*phys_mask = dma_to_phys(dev, dma_mask);
66 
67 	/*
68 	 * Optimistically try the zone that the physical address mask falls
69 	 * into first.  If that returns memory that isn't actually addressable
70 	 * we will fallback to the next lower zone and try again.
71 	 *
72 	 * Note that GFP_DMA32 and GFP_DMA are no ops without the corresponding
73 	 * zones.
74 	 */
75 	if (*phys_mask <= DMA_BIT_MASK(ARCH_ZONE_DMA_BITS))
76 		return GFP_DMA;
77 	if (*phys_mask <= DMA_BIT_MASK(32))
78 		return GFP_DMA32;
79 	return 0;
80 }
81 
82 static bool dma_coherent_ok(struct device *dev, phys_addr_t phys, size_t size)
83 {
84 	return phys_to_dma_direct(dev, phys) + size - 1 <=
85 			min_not_zero(dev->coherent_dma_mask, dev->bus_dma_mask);
86 }
87 
88 struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
89 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
90 {
91 	struct page *page = NULL;
92 	u64 phys_mask;
93 
94 	if (attrs & DMA_ATTR_NO_WARN)
95 		gfp |= __GFP_NOWARN;
96 
97 	/* we always manually zero the memory once we are done: */
98 	gfp &= ~__GFP_ZERO;
99 	gfp |= __dma_direct_optimal_gfp_mask(dev, dev->coherent_dma_mask,
100 			&phys_mask);
101 again:
102 	page = dma_alloc_contiguous(dev, size, gfp);
103 	if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
104 		dma_free_contiguous(dev, page, size);
105 		page = NULL;
106 
107 		if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
108 		    phys_mask < DMA_BIT_MASK(64) &&
109 		    !(gfp & (GFP_DMA32 | GFP_DMA))) {
110 			gfp |= GFP_DMA32;
111 			goto again;
112 		}
113 
114 		if (IS_ENABLED(CONFIG_ZONE_DMA) && !(gfp & GFP_DMA)) {
115 			gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
116 			goto again;
117 		}
118 	}
119 
120 	return page;
121 }
122 
123 void *dma_direct_alloc_pages(struct device *dev, size_t size,
124 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
125 {
126 	struct page *page;
127 	void *ret;
128 
129 	page = __dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
130 	if (!page)
131 		return NULL;
132 
133 	if (attrs & DMA_ATTR_NO_KERNEL_MAPPING) {
134 		/* remove any dirty cache lines on the kernel alias */
135 		if (!PageHighMem(page))
136 			arch_dma_prep_coherent(page, size);
137 		/* return the page pointer as the opaque cookie */
138 		return page;
139 	}
140 
141 	if (PageHighMem(page)) {
142 		/*
143 		 * Depending on the cma= arguments and per-arch setup
144 		 * dma_alloc_contiguous could return highmem pages.
145 		 * Without remapping there is no way to return them here,
146 		 * so log an error and fail.
147 		 */
148 		dev_info(dev, "Rejecting highmem page from CMA.\n");
149 		__dma_direct_free_pages(dev, size, page);
150 		return NULL;
151 	}
152 
153 	ret = page_address(page);
154 	if (force_dma_unencrypted(dev)) {
155 		set_memory_decrypted((unsigned long)ret, 1 << get_order(size));
156 		*dma_handle = __phys_to_dma(dev, page_to_phys(page));
157 	} else {
158 		*dma_handle = phys_to_dma(dev, page_to_phys(page));
159 	}
160 	memset(ret, 0, size);
161 
162 	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
163 	    dma_alloc_need_uncached(dev, attrs)) {
164 		arch_dma_prep_coherent(page, size);
165 		ret = uncached_kernel_address(ret);
166 	}
167 
168 	return ret;
169 }
170 
171 void __dma_direct_free_pages(struct device *dev, size_t size, struct page *page)
172 {
173 	dma_free_contiguous(dev, page, size);
174 }
175 
176 void dma_direct_free_pages(struct device *dev, size_t size, void *cpu_addr,
177 		dma_addr_t dma_addr, unsigned long attrs)
178 {
179 	unsigned int page_order = get_order(size);
180 
181 	if (attrs & DMA_ATTR_NO_KERNEL_MAPPING) {
182 		/* cpu_addr is a struct page cookie, not a kernel address */
183 		__dma_direct_free_pages(dev, size, cpu_addr);
184 		return;
185 	}
186 
187 	if (force_dma_unencrypted(dev))
188 		set_memory_encrypted((unsigned long)cpu_addr, 1 << page_order);
189 
190 	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
191 	    dma_alloc_need_uncached(dev, attrs))
192 		cpu_addr = cached_kernel_address(cpu_addr);
193 	__dma_direct_free_pages(dev, size, virt_to_page(cpu_addr));
194 }
195 
196 void *dma_direct_alloc(struct device *dev, size_t size,
197 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
198 {
199 	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
200 	    dma_alloc_need_uncached(dev, attrs))
201 		return arch_dma_alloc(dev, size, dma_handle, gfp, attrs);
202 	return dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
203 }
204 
205 void dma_direct_free(struct device *dev, size_t size,
206 		void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs)
207 {
208 	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
209 	    dma_alloc_need_uncached(dev, attrs))
210 		arch_dma_free(dev, size, cpu_addr, dma_addr, attrs);
211 	else
212 		dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
213 }
214 
215 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_DEVICE) || \
216     defined(CONFIG_SWIOTLB)
217 void dma_direct_sync_single_for_device(struct device *dev,
218 		dma_addr_t addr, size_t size, enum dma_data_direction dir)
219 {
220 	phys_addr_t paddr = dma_to_phys(dev, addr);
221 
222 	if (unlikely(is_swiotlb_buffer(paddr)))
223 		swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_DEVICE);
224 
225 	if (!dev_is_dma_coherent(dev))
226 		arch_sync_dma_for_device(dev, paddr, size, dir);
227 }
228 EXPORT_SYMBOL(dma_direct_sync_single_for_device);
229 
230 void dma_direct_sync_sg_for_device(struct device *dev,
231 		struct scatterlist *sgl, int nents, enum dma_data_direction dir)
232 {
233 	struct scatterlist *sg;
234 	int i;
235 
236 	for_each_sg(sgl, sg, nents, i) {
237 		phys_addr_t paddr = dma_to_phys(dev, sg_dma_address(sg));
238 
239 		if (unlikely(is_swiotlb_buffer(paddr)))
240 			swiotlb_tbl_sync_single(dev, paddr, sg->length,
241 					dir, SYNC_FOR_DEVICE);
242 
243 		if (!dev_is_dma_coherent(dev))
244 			arch_sync_dma_for_device(dev, paddr, sg->length,
245 					dir);
246 	}
247 }
248 EXPORT_SYMBOL(dma_direct_sync_sg_for_device);
249 #endif
250 
251 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU) || \
252     defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_ALL) || \
253     defined(CONFIG_SWIOTLB)
254 void dma_direct_sync_single_for_cpu(struct device *dev,
255 		dma_addr_t addr, size_t size, enum dma_data_direction dir)
256 {
257 	phys_addr_t paddr = dma_to_phys(dev, addr);
258 
259 	if (!dev_is_dma_coherent(dev)) {
260 		arch_sync_dma_for_cpu(dev, paddr, size, dir);
261 		arch_sync_dma_for_cpu_all(dev);
262 	}
263 
264 	if (unlikely(is_swiotlb_buffer(paddr)))
265 		swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_CPU);
266 }
267 EXPORT_SYMBOL(dma_direct_sync_single_for_cpu);
268 
269 void dma_direct_sync_sg_for_cpu(struct device *dev,
270 		struct scatterlist *sgl, int nents, enum dma_data_direction dir)
271 {
272 	struct scatterlist *sg;
273 	int i;
274 
275 	for_each_sg(sgl, sg, nents, i) {
276 		phys_addr_t paddr = dma_to_phys(dev, sg_dma_address(sg));
277 
278 		if (!dev_is_dma_coherent(dev))
279 			arch_sync_dma_for_cpu(dev, paddr, sg->length, dir);
280 
281 		if (unlikely(is_swiotlb_buffer(paddr)))
282 			swiotlb_tbl_sync_single(dev, paddr, sg->length, dir,
283 					SYNC_FOR_CPU);
284 	}
285 
286 	if (!dev_is_dma_coherent(dev))
287 		arch_sync_dma_for_cpu_all(dev);
288 }
289 EXPORT_SYMBOL(dma_direct_sync_sg_for_cpu);
290 
291 void dma_direct_unmap_page(struct device *dev, dma_addr_t addr,
292 		size_t size, enum dma_data_direction dir, unsigned long attrs)
293 {
294 	phys_addr_t phys = dma_to_phys(dev, addr);
295 
296 	if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
297 		dma_direct_sync_single_for_cpu(dev, addr, size, dir);
298 
299 	if (unlikely(is_swiotlb_buffer(phys)))
300 		swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
301 }
302 EXPORT_SYMBOL(dma_direct_unmap_page);
303 
304 void dma_direct_unmap_sg(struct device *dev, struct scatterlist *sgl,
305 		int nents, enum dma_data_direction dir, unsigned long attrs)
306 {
307 	struct scatterlist *sg;
308 	int i;
309 
310 	for_each_sg(sgl, sg, nents, i)
311 		dma_direct_unmap_page(dev, sg->dma_address, sg_dma_len(sg), dir,
312 			     attrs);
313 }
314 EXPORT_SYMBOL(dma_direct_unmap_sg);
315 #endif
316 
317 static inline bool dma_direct_possible(struct device *dev, dma_addr_t dma_addr,
318 		size_t size)
319 {
320 	return swiotlb_force != SWIOTLB_FORCE &&
321 		dma_capable(dev, dma_addr, size);
322 }
323 
324 dma_addr_t dma_direct_map_page(struct device *dev, struct page *page,
325 		unsigned long offset, size_t size, enum dma_data_direction dir,
326 		unsigned long attrs)
327 {
328 	phys_addr_t phys = page_to_phys(page) + offset;
329 	dma_addr_t dma_addr = phys_to_dma(dev, phys);
330 
331 	if (unlikely(!dma_direct_possible(dev, dma_addr, size)) &&
332 	    !swiotlb_map(dev, &phys, &dma_addr, size, dir, attrs)) {
333 		report_addr(dev, dma_addr, size);
334 		return DMA_MAPPING_ERROR;
335 	}
336 
337 	if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
338 		arch_sync_dma_for_device(dev, phys, size, dir);
339 	return dma_addr;
340 }
341 EXPORT_SYMBOL(dma_direct_map_page);
342 
343 int dma_direct_map_sg(struct device *dev, struct scatterlist *sgl, int nents,
344 		enum dma_data_direction dir, unsigned long attrs)
345 {
346 	int i;
347 	struct scatterlist *sg;
348 
349 	for_each_sg(sgl, sg, nents, i) {
350 		sg->dma_address = dma_direct_map_page(dev, sg_page(sg),
351 				sg->offset, sg->length, dir, attrs);
352 		if (sg->dma_address == DMA_MAPPING_ERROR)
353 			goto out_unmap;
354 		sg_dma_len(sg) = sg->length;
355 	}
356 
357 	return nents;
358 
359 out_unmap:
360 	dma_direct_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
361 	return 0;
362 }
363 EXPORT_SYMBOL(dma_direct_map_sg);
364 
365 dma_addr_t dma_direct_map_resource(struct device *dev, phys_addr_t paddr,
366 		size_t size, enum dma_data_direction dir, unsigned long attrs)
367 {
368 	dma_addr_t dma_addr = paddr;
369 
370 	if (unlikely(!dma_direct_possible(dev, dma_addr, size))) {
371 		report_addr(dev, dma_addr, size);
372 		return DMA_MAPPING_ERROR;
373 	}
374 
375 	return dma_addr;
376 }
377 EXPORT_SYMBOL(dma_direct_map_resource);
378 
379 /*
380  * Because 32-bit DMA masks are so common we expect every architecture to be
381  * able to satisfy them - either by not supporting more physical memory, or by
382  * providing a ZONE_DMA32.  If neither is the case, the architecture needs to
383  * use an IOMMU instead of the direct mapping.
384  */
385 int dma_direct_supported(struct device *dev, u64 mask)
386 {
387 	u64 min_mask;
388 
389 	if (IS_ENABLED(CONFIG_ZONE_DMA))
390 		min_mask = DMA_BIT_MASK(ARCH_ZONE_DMA_BITS);
391 	else
392 		min_mask = DMA_BIT_MASK(32);
393 
394 	min_mask = min_t(u64, min_mask, (max_pfn - 1) << PAGE_SHIFT);
395 
396 	/*
397 	 * This check needs to be against the actual bit mask value, so
398 	 * use __phys_to_dma() here so that the SME encryption mask isn't
399 	 * part of the check.
400 	 */
401 	return mask >= __phys_to_dma(dev, min_mask);
402 }
403 
404 size_t dma_direct_max_mapping_size(struct device *dev)
405 {
406 	/* If SWIOTLB is active, use its maximum mapping size */
407 	if (is_swiotlb_active() &&
408 	    (dma_addressing_limited(dev) || swiotlb_force == SWIOTLB_FORCE))
409 		return swiotlb_max_mapping_size(dev);
410 	return SIZE_MAX;
411 }
412