xref: /openbmc/linux/kernel/dma/direct.c (revision d89775fc)
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/pfn.h>
14 #include <linux/vmalloc.h>
15 #include <linux/set_memory.h>
16 
17 /*
18  * Most architectures use ZONE_DMA for the first 16 Megabytes, but some use it
19  * it for entirely different regions. In that case the arch code needs to
20  * override the variable below for dma-direct to work properly.
21  */
22 unsigned int zone_dma_bits __ro_after_init = 24;
23 
24 static inline dma_addr_t phys_to_dma_direct(struct device *dev,
25 		phys_addr_t phys)
26 {
27 	if (force_dma_unencrypted(dev))
28 		return __phys_to_dma(dev, phys);
29 	return phys_to_dma(dev, phys);
30 }
31 
32 static inline struct page *dma_direct_to_page(struct device *dev,
33 		dma_addr_t dma_addr)
34 {
35 	return pfn_to_page(PHYS_PFN(dma_to_phys(dev, dma_addr)));
36 }
37 
38 u64 dma_direct_get_required_mask(struct device *dev)
39 {
40 	phys_addr_t phys = (phys_addr_t)(max_pfn - 1) << PAGE_SHIFT;
41 	u64 max_dma = phys_to_dma_direct(dev, phys);
42 
43 	return (1ULL << (fls64(max_dma) - 1)) * 2 - 1;
44 }
45 
46 gfp_t dma_direct_optimal_gfp_mask(struct device *dev, u64 dma_mask,
47 				  u64 *phys_limit)
48 {
49 	u64 dma_limit = min_not_zero(dma_mask, dev->bus_dma_limit);
50 
51 	if (force_dma_unencrypted(dev))
52 		*phys_limit = __dma_to_phys(dev, dma_limit);
53 	else
54 		*phys_limit = dma_to_phys(dev, dma_limit);
55 
56 	/*
57 	 * Optimistically try the zone that the physical address mask falls
58 	 * into first.  If that returns memory that isn't actually addressable
59 	 * we will fallback to the next lower zone and try again.
60 	 *
61 	 * Note that GFP_DMA32 and GFP_DMA are no ops without the corresponding
62 	 * zones.
63 	 */
64 	if (*phys_limit <= DMA_BIT_MASK(zone_dma_bits))
65 		return GFP_DMA;
66 	if (*phys_limit <= DMA_BIT_MASK(32))
67 		return GFP_DMA32;
68 	return 0;
69 }
70 
71 bool dma_coherent_ok(struct device *dev, phys_addr_t phys, size_t size)
72 {
73 	return phys_to_dma_direct(dev, phys) + size - 1 <=
74 			min_not_zero(dev->coherent_dma_mask, dev->bus_dma_limit);
75 }
76 
77 /*
78  * Decrypting memory is allowed to block, so if this device requires
79  * unencrypted memory it must come from atomic pools.
80  */
81 static inline bool dma_should_alloc_from_pool(struct device *dev, gfp_t gfp,
82 					      unsigned long attrs)
83 {
84 	if (!IS_ENABLED(CONFIG_DMA_COHERENT_POOL))
85 		return false;
86 	if (gfpflags_allow_blocking(gfp))
87 		return false;
88 	if (force_dma_unencrypted(dev))
89 		return true;
90 	if (!IS_ENABLED(CONFIG_DMA_DIRECT_REMAP))
91 		return false;
92 	if (dma_alloc_need_uncached(dev, attrs))
93 		return true;
94 	return false;
95 }
96 
97 static inline bool dma_should_free_from_pool(struct device *dev,
98 					     unsigned long attrs)
99 {
100 	if (IS_ENABLED(CONFIG_DMA_COHERENT_POOL))
101 		return true;
102 	if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) &&
103 	    !force_dma_unencrypted(dev))
104 		return false;
105 	if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP))
106 		return true;
107 	return false;
108 }
109 
110 static struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
111 		gfp_t gfp, unsigned long attrs)
112 {
113 	int node = dev_to_node(dev);
114 	struct page *page = NULL;
115 	u64 phys_limit;
116 
117 	WARN_ON_ONCE(!PAGE_ALIGNED(size));
118 
119 	if (attrs & DMA_ATTR_NO_WARN)
120 		gfp |= __GFP_NOWARN;
121 
122 	/* we always manually zero the memory once we are done: */
123 	gfp &= ~__GFP_ZERO;
124 	gfp |= dma_direct_optimal_gfp_mask(dev, dev->coherent_dma_mask,
125 					   &phys_limit);
126 	page = dma_alloc_contiguous(dev, size, gfp);
127 	if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
128 		dma_free_contiguous(dev, page, size);
129 		page = NULL;
130 	}
131 again:
132 	if (!page)
133 		page = alloc_pages_node(node, gfp, get_order(size));
134 	if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
135 		dma_free_contiguous(dev, page, size);
136 		page = NULL;
137 
138 		if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
139 		    phys_limit < DMA_BIT_MASK(64) &&
140 		    !(gfp & (GFP_DMA32 | GFP_DMA))) {
141 			gfp |= GFP_DMA32;
142 			goto again;
143 		}
144 
145 		if (IS_ENABLED(CONFIG_ZONE_DMA) && !(gfp & GFP_DMA)) {
146 			gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
147 			goto again;
148 		}
149 	}
150 
151 	return page;
152 }
153 
154 void *dma_direct_alloc_pages(struct device *dev, size_t size,
155 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
156 {
157 	struct page *page;
158 	void *ret;
159 	int err;
160 
161 	size = PAGE_ALIGN(size);
162 
163 	if (dma_should_alloc_from_pool(dev, gfp, attrs)) {
164 		ret = dma_alloc_from_pool(dev, size, &page, gfp);
165 		if (!ret)
166 			return NULL;
167 		goto done;
168 	}
169 
170 	page = __dma_direct_alloc_pages(dev, size, gfp, attrs);
171 	if (!page)
172 		return NULL;
173 
174 	if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) &&
175 	    !force_dma_unencrypted(dev)) {
176 		/* remove any dirty cache lines on the kernel alias */
177 		if (!PageHighMem(page))
178 			arch_dma_prep_coherent(page, size);
179 		/* return the page pointer as the opaque cookie */
180 		ret = page;
181 		goto done;
182 	}
183 
184 	if ((IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
185 	     dma_alloc_need_uncached(dev, attrs)) ||
186 	    (IS_ENABLED(CONFIG_DMA_REMAP) && PageHighMem(page))) {
187 		/* remove any dirty cache lines on the kernel alias */
188 		arch_dma_prep_coherent(page, size);
189 
190 		/* create a coherent mapping */
191 		ret = dma_common_contiguous_remap(page, size,
192 				dma_pgprot(dev, PAGE_KERNEL, attrs),
193 				__builtin_return_address(0));
194 		if (!ret)
195 			goto out_free_pages;
196 		if (force_dma_unencrypted(dev)) {
197 			err = set_memory_decrypted((unsigned long)ret,
198 						   1 << get_order(size));
199 			if (err)
200 				goto out_free_pages;
201 		}
202 		memset(ret, 0, size);
203 		goto done;
204 	}
205 
206 	if (PageHighMem(page)) {
207 		/*
208 		 * Depending on the cma= arguments and per-arch setup
209 		 * dma_alloc_contiguous could return highmem pages.
210 		 * Without remapping there is no way to return them here,
211 		 * so log an error and fail.
212 		 */
213 		dev_info(dev, "Rejecting highmem page from CMA.\n");
214 		goto out_free_pages;
215 	}
216 
217 	ret = page_address(page);
218 	if (force_dma_unencrypted(dev)) {
219 		err = set_memory_decrypted((unsigned long)ret,
220 					   1 << get_order(size));
221 		if (err)
222 			goto out_free_pages;
223 	}
224 
225 	memset(ret, 0, size);
226 
227 	if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_SET_UNCACHED) &&
228 	    dma_alloc_need_uncached(dev, attrs)) {
229 		arch_dma_prep_coherent(page, size);
230 		ret = arch_dma_set_uncached(ret, size);
231 		if (IS_ERR(ret))
232 			goto out_encrypt_pages;
233 	}
234 done:
235 	if (force_dma_unencrypted(dev))
236 		*dma_handle = __phys_to_dma(dev, page_to_phys(page));
237 	else
238 		*dma_handle = phys_to_dma(dev, page_to_phys(page));
239 	return ret;
240 
241 out_encrypt_pages:
242 	if (force_dma_unencrypted(dev)) {
243 		err = set_memory_encrypted((unsigned long)page_address(page),
244 					   1 << get_order(size));
245 		/* If memory cannot be re-encrypted, it must be leaked */
246 		if (err)
247 			return NULL;
248 	}
249 out_free_pages:
250 	dma_free_contiguous(dev, page, size);
251 	return NULL;
252 }
253 
254 void dma_direct_free_pages(struct device *dev, size_t size, void *cpu_addr,
255 		dma_addr_t dma_addr, unsigned long attrs)
256 {
257 	unsigned int page_order = get_order(size);
258 
259 	/* If cpu_addr is not from an atomic pool, dma_free_from_pool() fails */
260 	if (dma_should_free_from_pool(dev, attrs) &&
261 	    dma_free_from_pool(dev, cpu_addr, PAGE_ALIGN(size)))
262 		return;
263 
264 	if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) &&
265 	    !force_dma_unencrypted(dev)) {
266 		/* cpu_addr is a struct page cookie, not a kernel address */
267 		dma_free_contiguous(dev, cpu_addr, size);
268 		return;
269 	}
270 
271 	if (force_dma_unencrypted(dev))
272 		set_memory_encrypted((unsigned long)cpu_addr, 1 << page_order);
273 
274 	if (IS_ENABLED(CONFIG_DMA_REMAP) && is_vmalloc_addr(cpu_addr))
275 		vunmap(cpu_addr);
276 	else if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_CLEAR_UNCACHED))
277 		arch_dma_clear_uncached(cpu_addr, size);
278 
279 	dma_free_contiguous(dev, dma_direct_to_page(dev, dma_addr), size);
280 }
281 
282 void *dma_direct_alloc(struct device *dev, size_t size,
283 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
284 {
285 	if (!IS_ENABLED(CONFIG_ARCH_HAS_DMA_SET_UNCACHED) &&
286 	    !IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
287 	    dma_alloc_need_uncached(dev, attrs))
288 		return arch_dma_alloc(dev, size, dma_handle, gfp, attrs);
289 	return dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
290 }
291 
292 void dma_direct_free(struct device *dev, size_t size,
293 		void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs)
294 {
295 	if (!IS_ENABLED(CONFIG_ARCH_HAS_DMA_SET_UNCACHED) &&
296 	    !IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
297 	    dma_alloc_need_uncached(dev, attrs))
298 		arch_dma_free(dev, size, cpu_addr, dma_addr, attrs);
299 	else
300 		dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
301 }
302 
303 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_DEVICE) || \
304     defined(CONFIG_SWIOTLB)
305 void dma_direct_sync_sg_for_device(struct device *dev,
306 		struct scatterlist *sgl, int nents, enum dma_data_direction dir)
307 {
308 	struct scatterlist *sg;
309 	int i;
310 
311 	for_each_sg(sgl, sg, nents, i) {
312 		phys_addr_t paddr = dma_to_phys(dev, sg_dma_address(sg));
313 
314 		if (unlikely(is_swiotlb_buffer(paddr)))
315 			swiotlb_tbl_sync_single(dev, paddr, sg->length,
316 					dir, SYNC_FOR_DEVICE);
317 
318 		if (!dev_is_dma_coherent(dev))
319 			arch_sync_dma_for_device(paddr, sg->length,
320 					dir);
321 	}
322 }
323 #endif
324 
325 #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU) || \
326     defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_ALL) || \
327     defined(CONFIG_SWIOTLB)
328 void dma_direct_sync_sg_for_cpu(struct device *dev,
329 		struct scatterlist *sgl, int nents, enum dma_data_direction dir)
330 {
331 	struct scatterlist *sg;
332 	int i;
333 
334 	for_each_sg(sgl, sg, nents, i) {
335 		phys_addr_t paddr = dma_to_phys(dev, sg_dma_address(sg));
336 
337 		if (!dev_is_dma_coherent(dev))
338 			arch_sync_dma_for_cpu(paddr, sg->length, dir);
339 
340 		if (unlikely(is_swiotlb_buffer(paddr)))
341 			swiotlb_tbl_sync_single(dev, paddr, sg->length, dir,
342 					SYNC_FOR_CPU);
343 	}
344 
345 	if (!dev_is_dma_coherent(dev))
346 		arch_sync_dma_for_cpu_all();
347 }
348 
349 void dma_direct_unmap_sg(struct device *dev, struct scatterlist *sgl,
350 		int nents, enum dma_data_direction dir, unsigned long attrs)
351 {
352 	struct scatterlist *sg;
353 	int i;
354 
355 	for_each_sg(sgl, sg, nents, i)
356 		dma_direct_unmap_page(dev, sg->dma_address, sg_dma_len(sg), dir,
357 			     attrs);
358 }
359 #endif
360 
361 int dma_direct_map_sg(struct device *dev, struct scatterlist *sgl, int nents,
362 		enum dma_data_direction dir, unsigned long attrs)
363 {
364 	int i;
365 	struct scatterlist *sg;
366 
367 	for_each_sg(sgl, sg, nents, i) {
368 		sg->dma_address = dma_direct_map_page(dev, sg_page(sg),
369 				sg->offset, sg->length, dir, attrs);
370 		if (sg->dma_address == DMA_MAPPING_ERROR)
371 			goto out_unmap;
372 		sg_dma_len(sg) = sg->length;
373 	}
374 
375 	return nents;
376 
377 out_unmap:
378 	dma_direct_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
379 	return 0;
380 }
381 
382 dma_addr_t dma_direct_map_resource(struct device *dev, phys_addr_t paddr,
383 		size_t size, enum dma_data_direction dir, unsigned long attrs)
384 {
385 	dma_addr_t dma_addr = paddr;
386 
387 	if (unlikely(!dma_capable(dev, dma_addr, size, false))) {
388 		dev_err_once(dev,
389 			     "DMA addr %pad+%zu overflow (mask %llx, bus limit %llx).\n",
390 			     &dma_addr, size, *dev->dma_mask, dev->bus_dma_limit);
391 		WARN_ON_ONCE(1);
392 		return DMA_MAPPING_ERROR;
393 	}
394 
395 	return dma_addr;
396 }
397 
398 int dma_direct_get_sgtable(struct device *dev, struct sg_table *sgt,
399 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
400 		unsigned long attrs)
401 {
402 	struct page *page = dma_direct_to_page(dev, dma_addr);
403 	int ret;
404 
405 	ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
406 	if (!ret)
407 		sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
408 	return ret;
409 }
410 
411 bool dma_direct_can_mmap(struct device *dev)
412 {
413 	return dev_is_dma_coherent(dev) ||
414 		IS_ENABLED(CONFIG_DMA_NONCOHERENT_MMAP);
415 }
416 
417 int dma_direct_mmap(struct device *dev, struct vm_area_struct *vma,
418 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
419 		unsigned long attrs)
420 {
421 	unsigned long user_count = vma_pages(vma);
422 	unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT;
423 	unsigned long pfn = PHYS_PFN(dma_to_phys(dev, dma_addr));
424 	int ret = -ENXIO;
425 
426 	vma->vm_page_prot = dma_pgprot(dev, vma->vm_page_prot, attrs);
427 
428 	if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
429 		return ret;
430 
431 	if (vma->vm_pgoff >= count || user_count > count - vma->vm_pgoff)
432 		return -ENXIO;
433 	return remap_pfn_range(vma, vma->vm_start, pfn + vma->vm_pgoff,
434 			user_count << PAGE_SHIFT, vma->vm_page_prot);
435 }
436 
437 int dma_direct_supported(struct device *dev, u64 mask)
438 {
439 	u64 min_mask = (max_pfn - 1) << PAGE_SHIFT;
440 
441 	/*
442 	 * Because 32-bit DMA masks are so common we expect every architecture
443 	 * to be able to satisfy them - either by not supporting more physical
444 	 * memory, or by providing a ZONE_DMA32.  If neither is the case, the
445 	 * architecture needs to use an IOMMU instead of the direct mapping.
446 	 */
447 	if (mask >= DMA_BIT_MASK(32))
448 		return 1;
449 
450 	/*
451 	 * This check needs to be against the actual bit mask value, so
452 	 * use __phys_to_dma() here so that the SME encryption mask isn't
453 	 * part of the check.
454 	 */
455 	if (IS_ENABLED(CONFIG_ZONE_DMA))
456 		min_mask = min_t(u64, min_mask, DMA_BIT_MASK(zone_dma_bits));
457 	return mask >= __phys_to_dma(dev, min_mask);
458 }
459 
460 size_t dma_direct_max_mapping_size(struct device *dev)
461 {
462 	/* If SWIOTLB is active, use its maximum mapping size */
463 	if (is_swiotlb_active() &&
464 	    (dma_addressing_limited(dev) || swiotlb_force == SWIOTLB_FORCE))
465 		return swiotlb_max_mapping_size(dev);
466 	return SIZE_MAX;
467 }
468 
469 bool dma_direct_need_sync(struct device *dev, dma_addr_t dma_addr)
470 {
471 	return !dev_is_dma_coherent(dev) ||
472 		is_swiotlb_buffer(dma_to_phys(dev, dma_addr));
473 }
474