xref: /openbmc/linux/kernel/dma/direct.c (revision 388f6966)
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/vmalloc.h>
16 #include <linux/set_memory.h>
17 #include <linux/swiotlb.h>
18 
19 /*
20  * Most architectures use ZONE_DMA for the first 16 Megabytes, but some use it
21  * it for entirely different regions. In that case the arch code needs to
22  * override the variable below for dma-direct to work properly.
23  */
24 unsigned int zone_dma_bits __ro_after_init = 24;
25 
26 static inline dma_addr_t phys_to_dma_direct(struct device *dev,
27 		phys_addr_t phys)
28 {
29 	if (force_dma_unencrypted(dev))
30 		return __phys_to_dma(dev, phys);
31 	return phys_to_dma(dev, phys);
32 }
33 
34 static inline struct page *dma_direct_to_page(struct device *dev,
35 		dma_addr_t dma_addr)
36 {
37 	return pfn_to_page(PHYS_PFN(dma_to_phys(dev, dma_addr)));
38 }
39 
40 u64 dma_direct_get_required_mask(struct device *dev)
41 {
42 	u64 max_dma = phys_to_dma_direct(dev, (max_pfn - 1) << PAGE_SHIFT);
43 
44 	return (1ULL << (fls64(max_dma) - 1)) * 2 - 1;
45 }
46 
47 static gfp_t __dma_direct_optimal_gfp_mask(struct device *dev, u64 dma_mask,
48 		u64 *phys_limit)
49 {
50 	u64 dma_limit = min_not_zero(dma_mask, dev->bus_dma_limit);
51 
52 	if (force_dma_unencrypted(dev))
53 		*phys_limit = __dma_to_phys(dev, dma_limit);
54 	else
55 		*phys_limit = dma_to_phys(dev, dma_limit);
56 
57 	/*
58 	 * Optimistically try the zone that the physical address mask falls
59 	 * into first.  If that returns memory that isn't actually addressable
60 	 * we will fallback to the next lower zone and try again.
61 	 *
62 	 * Note that GFP_DMA32 and GFP_DMA are no ops without the corresponding
63 	 * zones.
64 	 */
65 	if (*phys_limit <= DMA_BIT_MASK(zone_dma_bits))
66 		return GFP_DMA;
67 	if (*phys_limit <= DMA_BIT_MASK(32))
68 		return GFP_DMA32;
69 	return 0;
70 }
71 
72 static bool dma_coherent_ok(struct device *dev, phys_addr_t phys, size_t size)
73 {
74 	return phys_to_dma_direct(dev, phys) + size - 1 <=
75 			min_not_zero(dev->coherent_dma_mask, dev->bus_dma_limit);
76 }
77 
78 struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
79 		gfp_t gfp, unsigned long attrs)
80 {
81 	size_t alloc_size = PAGE_ALIGN(size);
82 	int node = dev_to_node(dev);
83 	struct page *page = NULL;
84 	u64 phys_limit;
85 
86 	if (attrs & DMA_ATTR_NO_WARN)
87 		gfp |= __GFP_NOWARN;
88 
89 	/* we always manually zero the memory once we are done: */
90 	gfp &= ~__GFP_ZERO;
91 	gfp |= __dma_direct_optimal_gfp_mask(dev, dev->coherent_dma_mask,
92 			&phys_limit);
93 	page = dma_alloc_contiguous(dev, alloc_size, gfp);
94 	if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
95 		dma_free_contiguous(dev, page, alloc_size);
96 		page = NULL;
97 	}
98 again:
99 	if (!page)
100 		page = alloc_pages_node(node, gfp, get_order(alloc_size));
101 	if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
102 		dma_free_contiguous(dev, page, size);
103 		page = NULL;
104 
105 		if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
106 		    phys_limit < DMA_BIT_MASK(64) &&
107 		    !(gfp & (GFP_DMA32 | GFP_DMA))) {
108 			gfp |= GFP_DMA32;
109 			goto again;
110 		}
111 
112 		if (IS_ENABLED(CONFIG_ZONE_DMA) && !(gfp & GFP_DMA)) {
113 			gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
114 			goto again;
115 		}
116 	}
117 
118 	return page;
119 }
120 
121 void *dma_direct_alloc_pages(struct device *dev, size_t size,
122 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
123 {
124 	struct page *page;
125 	void *ret;
126 
127 	if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
128 	    dma_alloc_need_uncached(dev, attrs) &&
129 	    !gfpflags_allow_blocking(gfp)) {
130 		ret = dma_alloc_from_pool(PAGE_ALIGN(size), &page, gfp);
131 		if (!ret)
132 			return NULL;
133 		goto done;
134 	}
135 
136 	page = __dma_direct_alloc_pages(dev, size, gfp, attrs);
137 	if (!page)
138 		return NULL;
139 
140 	if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) &&
141 	    !force_dma_unencrypted(dev)) {
142 		/* remove any dirty cache lines on the kernel alias */
143 		if (!PageHighMem(page))
144 			arch_dma_prep_coherent(page, size);
145 		/* return the page pointer as the opaque cookie */
146 		ret = page;
147 		goto done;
148 	}
149 
150 	if ((IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
151 	     dma_alloc_need_uncached(dev, attrs)) ||
152 	    (IS_ENABLED(CONFIG_DMA_REMAP) && PageHighMem(page))) {
153 		/* remove any dirty cache lines on the kernel alias */
154 		arch_dma_prep_coherent(page, PAGE_ALIGN(size));
155 
156 		/* create a coherent mapping */
157 		ret = dma_common_contiguous_remap(page, PAGE_ALIGN(size),
158 				dma_pgprot(dev, PAGE_KERNEL, attrs),
159 				__builtin_return_address(0));
160 		if (!ret) {
161 			dma_free_contiguous(dev, page, size);
162 			return ret;
163 		}
164 
165 		memset(ret, 0, size);
166 		goto done;
167 	}
168 
169 	if (PageHighMem(page)) {
170 		/*
171 		 * Depending on the cma= arguments and per-arch setup
172 		 * dma_alloc_contiguous could return highmem pages.
173 		 * Without remapping there is no way to return them here,
174 		 * so log an error and fail.
175 		 */
176 		dev_info(dev, "Rejecting highmem page from CMA.\n");
177 		dma_free_contiguous(dev, page, size);
178 		return NULL;
179 	}
180 
181 	ret = page_address(page);
182 	if (force_dma_unencrypted(dev))
183 		set_memory_decrypted((unsigned long)ret, 1 << get_order(size));
184 
185 	memset(ret, 0, size);
186 
187 	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
188 	    dma_alloc_need_uncached(dev, attrs)) {
189 		arch_dma_prep_coherent(page, size);
190 		ret = uncached_kernel_address(ret);
191 	}
192 done:
193 	if (force_dma_unencrypted(dev))
194 		*dma_handle = __phys_to_dma(dev, page_to_phys(page));
195 	else
196 		*dma_handle = phys_to_dma(dev, page_to_phys(page));
197 	return ret;
198 }
199 
200 void dma_direct_free_pages(struct device *dev, size_t size, void *cpu_addr,
201 		dma_addr_t dma_addr, unsigned long attrs)
202 {
203 	unsigned int page_order = get_order(size);
204 
205 	if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) &&
206 	    !force_dma_unencrypted(dev)) {
207 		/* cpu_addr is a struct page cookie, not a kernel address */
208 		dma_free_contiguous(dev, cpu_addr, size);
209 		return;
210 	}
211 
212 	if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
213 	    dma_free_from_pool(cpu_addr, PAGE_ALIGN(size)))
214 		return;
215 
216 	if (force_dma_unencrypted(dev))
217 		set_memory_encrypted((unsigned long)cpu_addr, 1 << page_order);
218 
219 	if (IS_ENABLED(CONFIG_DMA_REMAP) && is_vmalloc_addr(cpu_addr))
220 		vunmap(cpu_addr);
221 
222 	dma_free_contiguous(dev, dma_direct_to_page(dev, dma_addr), size);
223 }
224 
225 void *dma_direct_alloc(struct device *dev, size_t size,
226 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
227 {
228 	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
229 	    !IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
230 	    dma_alloc_need_uncached(dev, attrs))
231 		return arch_dma_alloc(dev, size, dma_handle, gfp, attrs);
232 	return dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
233 }
234 
235 void dma_direct_free(struct device *dev, size_t size,
236 		void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs)
237 {
238 	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
239 	    !IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
240 	    dma_alloc_need_uncached(dev, attrs))
241 		arch_dma_free(dev, size, cpu_addr, dma_addr, attrs);
242 	else
243 		dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
244 }
245 
246 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_DEVICE) || \
247     defined(CONFIG_SWIOTLB)
248 void dma_direct_sync_single_for_device(struct device *dev,
249 		dma_addr_t addr, size_t size, enum dma_data_direction dir)
250 {
251 	phys_addr_t paddr = dma_to_phys(dev, addr);
252 
253 	if (unlikely(is_swiotlb_buffer(paddr)))
254 		swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_DEVICE);
255 
256 	if (!dev_is_dma_coherent(dev))
257 		arch_sync_dma_for_device(paddr, size, dir);
258 }
259 EXPORT_SYMBOL(dma_direct_sync_single_for_device);
260 
261 void dma_direct_sync_sg_for_device(struct device *dev,
262 		struct scatterlist *sgl, int nents, enum dma_data_direction dir)
263 {
264 	struct scatterlist *sg;
265 	int i;
266 
267 	for_each_sg(sgl, sg, nents, i) {
268 		phys_addr_t paddr = dma_to_phys(dev, sg_dma_address(sg));
269 
270 		if (unlikely(is_swiotlb_buffer(paddr)))
271 			swiotlb_tbl_sync_single(dev, paddr, sg->length,
272 					dir, SYNC_FOR_DEVICE);
273 
274 		if (!dev_is_dma_coherent(dev))
275 			arch_sync_dma_for_device(paddr, sg->length,
276 					dir);
277 	}
278 }
279 EXPORT_SYMBOL(dma_direct_sync_sg_for_device);
280 #endif
281 
282 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU) || \
283     defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_ALL) || \
284     defined(CONFIG_SWIOTLB)
285 void dma_direct_sync_single_for_cpu(struct device *dev,
286 		dma_addr_t addr, size_t size, enum dma_data_direction dir)
287 {
288 	phys_addr_t paddr = dma_to_phys(dev, addr);
289 
290 	if (!dev_is_dma_coherent(dev)) {
291 		arch_sync_dma_for_cpu(paddr, size, dir);
292 		arch_sync_dma_for_cpu_all();
293 	}
294 
295 	if (unlikely(is_swiotlb_buffer(paddr)))
296 		swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_CPU);
297 }
298 EXPORT_SYMBOL(dma_direct_sync_single_for_cpu);
299 
300 void dma_direct_sync_sg_for_cpu(struct device *dev,
301 		struct scatterlist *sgl, int nents, enum dma_data_direction dir)
302 {
303 	struct scatterlist *sg;
304 	int i;
305 
306 	for_each_sg(sgl, sg, nents, i) {
307 		phys_addr_t paddr = dma_to_phys(dev, sg_dma_address(sg));
308 
309 		if (!dev_is_dma_coherent(dev))
310 			arch_sync_dma_for_cpu(paddr, sg->length, dir);
311 
312 		if (unlikely(is_swiotlb_buffer(paddr)))
313 			swiotlb_tbl_sync_single(dev, paddr, sg->length, dir,
314 					SYNC_FOR_CPU);
315 	}
316 
317 	if (!dev_is_dma_coherent(dev))
318 		arch_sync_dma_for_cpu_all();
319 }
320 EXPORT_SYMBOL(dma_direct_sync_sg_for_cpu);
321 
322 void dma_direct_unmap_page(struct device *dev, dma_addr_t addr,
323 		size_t size, enum dma_data_direction dir, unsigned long attrs)
324 {
325 	phys_addr_t phys = dma_to_phys(dev, addr);
326 
327 	if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
328 		dma_direct_sync_single_for_cpu(dev, addr, size, dir);
329 
330 	if (unlikely(is_swiotlb_buffer(phys)))
331 		swiotlb_tbl_unmap_single(dev, phys, size, size, dir, attrs);
332 }
333 EXPORT_SYMBOL(dma_direct_unmap_page);
334 
335 void dma_direct_unmap_sg(struct device *dev, struct scatterlist *sgl,
336 		int nents, enum dma_data_direction dir, unsigned long attrs)
337 {
338 	struct scatterlist *sg;
339 	int i;
340 
341 	for_each_sg(sgl, sg, nents, i)
342 		dma_direct_unmap_page(dev, sg->dma_address, sg_dma_len(sg), dir,
343 			     attrs);
344 }
345 EXPORT_SYMBOL(dma_direct_unmap_sg);
346 #endif
347 
348 dma_addr_t dma_direct_map_page(struct device *dev, struct page *page,
349 		unsigned long offset, size_t size, enum dma_data_direction dir,
350 		unsigned long attrs)
351 {
352 	phys_addr_t phys = page_to_phys(page) + offset;
353 	dma_addr_t dma_addr = phys_to_dma(dev, phys);
354 
355 	if (unlikely(swiotlb_force == SWIOTLB_FORCE))
356 		return swiotlb_map(dev, phys, size, dir, attrs);
357 
358 	if (unlikely(!dma_capable(dev, dma_addr, size, true))) {
359 		if (swiotlb_force != SWIOTLB_NO_FORCE)
360 			return swiotlb_map(dev, phys, size, dir, attrs);
361 
362 		dev_WARN_ONCE(dev, 1,
363 			     "DMA addr %pad+%zu overflow (mask %llx, bus limit %llx).\n",
364 			     &dma_addr, size, *dev->dma_mask, dev->bus_dma_limit);
365 		return DMA_MAPPING_ERROR;
366 	}
367 
368 	if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
369 		arch_sync_dma_for_device(phys, size, dir);
370 	return dma_addr;
371 }
372 EXPORT_SYMBOL(dma_direct_map_page);
373 
374 int dma_direct_map_sg(struct device *dev, struct scatterlist *sgl, int nents,
375 		enum dma_data_direction dir, unsigned long attrs)
376 {
377 	int i;
378 	struct scatterlist *sg;
379 
380 	for_each_sg(sgl, sg, nents, i) {
381 		sg->dma_address = dma_direct_map_page(dev, sg_page(sg),
382 				sg->offset, sg->length, dir, attrs);
383 		if (sg->dma_address == DMA_MAPPING_ERROR)
384 			goto out_unmap;
385 		sg_dma_len(sg) = sg->length;
386 	}
387 
388 	return nents;
389 
390 out_unmap:
391 	dma_direct_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
392 	return 0;
393 }
394 EXPORT_SYMBOL(dma_direct_map_sg);
395 
396 dma_addr_t dma_direct_map_resource(struct device *dev, phys_addr_t paddr,
397 		size_t size, enum dma_data_direction dir, unsigned long attrs)
398 {
399 	dma_addr_t dma_addr = paddr;
400 
401 	if (unlikely(!dma_capable(dev, dma_addr, size, false))) {
402 		dev_err_once(dev,
403 			     "DMA addr %pad+%zu overflow (mask %llx, bus limit %llx).\n",
404 			     &dma_addr, size, *dev->dma_mask, dev->bus_dma_limit);
405 		WARN_ON_ONCE(1);
406 		return DMA_MAPPING_ERROR;
407 	}
408 
409 	return dma_addr;
410 }
411 EXPORT_SYMBOL(dma_direct_map_resource);
412 
413 int dma_direct_get_sgtable(struct device *dev, struct sg_table *sgt,
414 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
415 		unsigned long attrs)
416 {
417 	struct page *page = dma_direct_to_page(dev, dma_addr);
418 	int ret;
419 
420 	ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
421 	if (!ret)
422 		sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
423 	return ret;
424 }
425 
426 #ifdef CONFIG_MMU
427 bool dma_direct_can_mmap(struct device *dev)
428 {
429 	return dev_is_dma_coherent(dev) ||
430 		IS_ENABLED(CONFIG_DMA_NONCOHERENT_MMAP);
431 }
432 
433 int dma_direct_mmap(struct device *dev, struct vm_area_struct *vma,
434 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
435 		unsigned long attrs)
436 {
437 	unsigned long user_count = vma_pages(vma);
438 	unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT;
439 	unsigned long pfn = PHYS_PFN(dma_to_phys(dev, dma_addr));
440 	int ret = -ENXIO;
441 
442 	vma->vm_page_prot = dma_pgprot(dev, vma->vm_page_prot, attrs);
443 
444 	if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
445 		return ret;
446 
447 	if (vma->vm_pgoff >= count || user_count > count - vma->vm_pgoff)
448 		return -ENXIO;
449 	return remap_pfn_range(vma, vma->vm_start, pfn + vma->vm_pgoff,
450 			user_count << PAGE_SHIFT, vma->vm_page_prot);
451 }
452 #else /* CONFIG_MMU */
453 bool dma_direct_can_mmap(struct device *dev)
454 {
455 	return false;
456 }
457 
458 int dma_direct_mmap(struct device *dev, struct vm_area_struct *vma,
459 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
460 		unsigned long attrs)
461 {
462 	return -ENXIO;
463 }
464 #endif /* CONFIG_MMU */
465 
466 int dma_direct_supported(struct device *dev, u64 mask)
467 {
468 	u64 min_mask = (max_pfn - 1) << PAGE_SHIFT;
469 
470 	/*
471 	 * Because 32-bit DMA masks are so common we expect every architecture
472 	 * to be able to satisfy them - either by not supporting more physical
473 	 * memory, or by providing a ZONE_DMA32.  If neither is the case, the
474 	 * architecture needs to use an IOMMU instead of the direct mapping.
475 	 */
476 	if (mask >= DMA_BIT_MASK(32))
477 		return 1;
478 
479 	/*
480 	 * This check needs to be against the actual bit mask value, so
481 	 * use __phys_to_dma() here so that the SME encryption mask isn't
482 	 * part of the check.
483 	 */
484 	if (IS_ENABLED(CONFIG_ZONE_DMA))
485 		min_mask = min_t(u64, min_mask, DMA_BIT_MASK(zone_dma_bits));
486 	return mask >= __phys_to_dma(dev, min_mask);
487 }
488 
489 size_t dma_direct_max_mapping_size(struct device *dev)
490 {
491 	/* If SWIOTLB is active, use its maximum mapping size */
492 	if (is_swiotlb_active() &&
493 	    (dma_addressing_limited(dev) || swiotlb_force == SWIOTLB_FORCE))
494 		return swiotlb_max_mapping_size(dev);
495 	return SIZE_MAX;
496 }
497