xref: /linux/kernel/dma/mapping.c (revision ed4bc1890b4984d0af447ad3cc1f93541623f8f3)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * arch-independent dma-mapping routines
4  *
5  * Copyright (c) 2006  SUSE Linux Products GmbH
6  * Copyright (c) 2006  Tejun Heo <teheo@suse.de>
7  */
8 #include <linux/memblock.h> /* for max_pfn */
9 #include <linux/acpi.h>
10 #include <linux/dma-direct.h>
11 #include <linux/dma-mapping.h>
12 #include <linux/export.h>
13 #include <linux/gfp.h>
14 #include <linux/of_device.h>
15 #include <linux/slab.h>
16 #include <linux/vmalloc.h>
17 
18 /*
19  * Managed DMA API
20  */
21 struct dma_devres {
22 	size_t		size;
23 	void		*vaddr;
24 	dma_addr_t	dma_handle;
25 	unsigned long	attrs;
26 };
27 
28 static void dmam_release(struct device *dev, void *res)
29 {
30 	struct dma_devres *this = res;
31 
32 	dma_free_attrs(dev, this->size, this->vaddr, this->dma_handle,
33 			this->attrs);
34 }
35 
36 static int dmam_match(struct device *dev, void *res, void *match_data)
37 {
38 	struct dma_devres *this = res, *match = match_data;
39 
40 	if (this->vaddr == match->vaddr) {
41 		WARN_ON(this->size != match->size ||
42 			this->dma_handle != match->dma_handle);
43 		return 1;
44 	}
45 	return 0;
46 }
47 
48 /**
49  * dmam_free_coherent - Managed dma_free_coherent()
50  * @dev: Device to free coherent memory for
51  * @size: Size of allocation
52  * @vaddr: Virtual address of the memory to free
53  * @dma_handle: DMA handle of the memory to free
54  *
55  * Managed dma_free_coherent().
56  */
57 void dmam_free_coherent(struct device *dev, size_t size, void *vaddr,
58 			dma_addr_t dma_handle)
59 {
60 	struct dma_devres match_data = { size, vaddr, dma_handle };
61 
62 	dma_free_coherent(dev, size, vaddr, dma_handle);
63 	WARN_ON(devres_destroy(dev, dmam_release, dmam_match, &match_data));
64 }
65 EXPORT_SYMBOL(dmam_free_coherent);
66 
67 /**
68  * dmam_alloc_attrs - Managed dma_alloc_attrs()
69  * @dev: Device to allocate non_coherent memory for
70  * @size: Size of allocation
71  * @dma_handle: Out argument for allocated DMA handle
72  * @gfp: Allocation flags
73  * @attrs: Flags in the DMA_ATTR_* namespace.
74  *
75  * Managed dma_alloc_attrs().  Memory allocated using this function will be
76  * automatically released on driver detach.
77  *
78  * RETURNS:
79  * Pointer to allocated memory on success, NULL on failure.
80  */
81 void *dmam_alloc_attrs(struct device *dev, size_t size, dma_addr_t *dma_handle,
82 		gfp_t gfp, unsigned long attrs)
83 {
84 	struct dma_devres *dr;
85 	void *vaddr;
86 
87 	dr = devres_alloc(dmam_release, sizeof(*dr), gfp);
88 	if (!dr)
89 		return NULL;
90 
91 	vaddr = dma_alloc_attrs(dev, size, dma_handle, gfp, attrs);
92 	if (!vaddr) {
93 		devres_free(dr);
94 		return NULL;
95 	}
96 
97 	dr->vaddr = vaddr;
98 	dr->dma_handle = *dma_handle;
99 	dr->size = size;
100 	dr->attrs = attrs;
101 
102 	devres_add(dev, dr);
103 
104 	return vaddr;
105 }
106 EXPORT_SYMBOL(dmam_alloc_attrs);
107 
108 static bool dma_go_direct(struct device *dev, dma_addr_t mask,
109 		const struct dma_map_ops *ops)
110 {
111 	if (likely(!ops))
112 		return true;
113 #ifdef CONFIG_DMA_OPS_BYPASS
114 	if (dev->dma_ops_bypass)
115 		return min_not_zero(mask, dev->bus_dma_limit) >=
116 			    dma_direct_get_required_mask(dev);
117 #endif
118 	return false;
119 }
120 
121 
122 /*
123  * Check if the devices uses a direct mapping for streaming DMA operations.
124  * This allows IOMMU drivers to set a bypass mode if the DMA mask is large
125  * enough.
126  */
127 static inline bool dma_alloc_direct(struct device *dev,
128 		const struct dma_map_ops *ops)
129 {
130 	return dma_go_direct(dev, dev->coherent_dma_mask, ops);
131 }
132 
133 static inline bool dma_map_direct(struct device *dev,
134 		const struct dma_map_ops *ops)
135 {
136 	return dma_go_direct(dev, *dev->dma_mask, ops);
137 }
138 
139 dma_addr_t dma_map_page_attrs(struct device *dev, struct page *page,
140 		size_t offset, size_t size, enum dma_data_direction dir,
141 		unsigned long attrs)
142 {
143 	const struct dma_map_ops *ops = get_dma_ops(dev);
144 	dma_addr_t addr;
145 
146 	BUG_ON(!valid_dma_direction(dir));
147 
148 	if (WARN_ON_ONCE(!dev->dma_mask))
149 		return DMA_MAPPING_ERROR;
150 
151 	if (dma_map_direct(dev, ops))
152 		addr = dma_direct_map_page(dev, page, offset, size, dir, attrs);
153 	else
154 		addr = ops->map_page(dev, page, offset, size, dir, attrs);
155 	debug_dma_map_page(dev, page, offset, size, dir, addr);
156 
157 	return addr;
158 }
159 EXPORT_SYMBOL(dma_map_page_attrs);
160 
161 void dma_unmap_page_attrs(struct device *dev, dma_addr_t addr, size_t size,
162 		enum dma_data_direction dir, unsigned long attrs)
163 {
164 	const struct dma_map_ops *ops = get_dma_ops(dev);
165 
166 	BUG_ON(!valid_dma_direction(dir));
167 	if (dma_map_direct(dev, ops))
168 		dma_direct_unmap_page(dev, addr, size, dir, attrs);
169 	else if (ops->unmap_page)
170 		ops->unmap_page(dev, addr, size, dir, attrs);
171 	debug_dma_unmap_page(dev, addr, size, dir);
172 }
173 EXPORT_SYMBOL(dma_unmap_page_attrs);
174 
175 /*
176  * dma_maps_sg_attrs returns 0 on error and > 0 on success.
177  * It should never return a value < 0.
178  */
179 int dma_map_sg_attrs(struct device *dev, struct scatterlist *sg, int nents,
180 		enum dma_data_direction dir, unsigned long attrs)
181 {
182 	const struct dma_map_ops *ops = get_dma_ops(dev);
183 	int ents;
184 
185 	BUG_ON(!valid_dma_direction(dir));
186 
187 	if (WARN_ON_ONCE(!dev->dma_mask))
188 		return 0;
189 
190 	if (dma_map_direct(dev, ops))
191 		ents = dma_direct_map_sg(dev, sg, nents, dir, attrs);
192 	else
193 		ents = ops->map_sg(dev, sg, nents, dir, attrs);
194 	BUG_ON(ents < 0);
195 	debug_dma_map_sg(dev, sg, nents, ents, dir);
196 
197 	return ents;
198 }
199 EXPORT_SYMBOL(dma_map_sg_attrs);
200 
201 void dma_unmap_sg_attrs(struct device *dev, struct scatterlist *sg,
202 				      int nents, enum dma_data_direction dir,
203 				      unsigned long attrs)
204 {
205 	const struct dma_map_ops *ops = get_dma_ops(dev);
206 
207 	BUG_ON(!valid_dma_direction(dir));
208 	debug_dma_unmap_sg(dev, sg, nents, dir);
209 	if (dma_map_direct(dev, ops))
210 		dma_direct_unmap_sg(dev, sg, nents, dir, attrs);
211 	else if (ops->unmap_sg)
212 		ops->unmap_sg(dev, sg, nents, dir, attrs);
213 }
214 EXPORT_SYMBOL(dma_unmap_sg_attrs);
215 
216 dma_addr_t dma_map_resource(struct device *dev, phys_addr_t phys_addr,
217 		size_t size, enum dma_data_direction dir, unsigned long attrs)
218 {
219 	const struct dma_map_ops *ops = get_dma_ops(dev);
220 	dma_addr_t addr = DMA_MAPPING_ERROR;
221 
222 	BUG_ON(!valid_dma_direction(dir));
223 
224 	if (WARN_ON_ONCE(!dev->dma_mask))
225 		return DMA_MAPPING_ERROR;
226 
227 	/* Don't allow RAM to be mapped */
228 	if (WARN_ON_ONCE(pfn_valid(PHYS_PFN(phys_addr))))
229 		return DMA_MAPPING_ERROR;
230 
231 	if (dma_map_direct(dev, ops))
232 		addr = dma_direct_map_resource(dev, phys_addr, size, dir, attrs);
233 	else if (ops->map_resource)
234 		addr = ops->map_resource(dev, phys_addr, size, dir, attrs);
235 
236 	debug_dma_map_resource(dev, phys_addr, size, dir, addr);
237 	return addr;
238 }
239 EXPORT_SYMBOL(dma_map_resource);
240 
241 void dma_unmap_resource(struct device *dev, dma_addr_t addr, size_t size,
242 		enum dma_data_direction dir, unsigned long attrs)
243 {
244 	const struct dma_map_ops *ops = get_dma_ops(dev);
245 
246 	BUG_ON(!valid_dma_direction(dir));
247 	if (!dma_map_direct(dev, ops) && ops->unmap_resource)
248 		ops->unmap_resource(dev, addr, size, dir, attrs);
249 	debug_dma_unmap_resource(dev, addr, size, dir);
250 }
251 EXPORT_SYMBOL(dma_unmap_resource);
252 
253 void dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
254 		enum dma_data_direction dir)
255 {
256 	const struct dma_map_ops *ops = get_dma_ops(dev);
257 
258 	BUG_ON(!valid_dma_direction(dir));
259 	if (dma_map_direct(dev, ops))
260 		dma_direct_sync_single_for_cpu(dev, addr, size, dir);
261 	else if (ops->sync_single_for_cpu)
262 		ops->sync_single_for_cpu(dev, addr, size, dir);
263 	debug_dma_sync_single_for_cpu(dev, addr, size, dir);
264 }
265 EXPORT_SYMBOL(dma_sync_single_for_cpu);
266 
267 void dma_sync_single_for_device(struct device *dev, dma_addr_t addr,
268 		size_t size, enum dma_data_direction dir)
269 {
270 	const struct dma_map_ops *ops = get_dma_ops(dev);
271 
272 	BUG_ON(!valid_dma_direction(dir));
273 	if (dma_map_direct(dev, ops))
274 		dma_direct_sync_single_for_device(dev, addr, size, dir);
275 	else if (ops->sync_single_for_device)
276 		ops->sync_single_for_device(dev, addr, size, dir);
277 	debug_dma_sync_single_for_device(dev, addr, size, dir);
278 }
279 EXPORT_SYMBOL(dma_sync_single_for_device);
280 
281 void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
282 		    int nelems, enum dma_data_direction dir)
283 {
284 	const struct dma_map_ops *ops = get_dma_ops(dev);
285 
286 	BUG_ON(!valid_dma_direction(dir));
287 	if (dma_map_direct(dev, ops))
288 		dma_direct_sync_sg_for_cpu(dev, sg, nelems, dir);
289 	else if (ops->sync_sg_for_cpu)
290 		ops->sync_sg_for_cpu(dev, sg, nelems, dir);
291 	debug_dma_sync_sg_for_cpu(dev, sg, nelems, dir);
292 }
293 EXPORT_SYMBOL(dma_sync_sg_for_cpu);
294 
295 void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
296 		       int nelems, enum dma_data_direction dir)
297 {
298 	const struct dma_map_ops *ops = get_dma_ops(dev);
299 
300 	BUG_ON(!valid_dma_direction(dir));
301 	if (dma_map_direct(dev, ops))
302 		dma_direct_sync_sg_for_device(dev, sg, nelems, dir);
303 	else if (ops->sync_sg_for_device)
304 		ops->sync_sg_for_device(dev, sg, nelems, dir);
305 	debug_dma_sync_sg_for_device(dev, sg, nelems, dir);
306 }
307 EXPORT_SYMBOL(dma_sync_sg_for_device);
308 
309 /*
310  * The whole dma_get_sgtable() idea is fundamentally unsafe - it seems
311  * that the intention is to allow exporting memory allocated via the
312  * coherent DMA APIs through the dma_buf API, which only accepts a
313  * scattertable.  This presents a couple of problems:
314  * 1. Not all memory allocated via the coherent DMA APIs is backed by
315  *    a struct page
316  * 2. Passing coherent DMA memory into the streaming APIs is not allowed
317  *    as we will try to flush the memory through a different alias to that
318  *    actually being used (and the flushes are redundant.)
319  */
320 int dma_get_sgtable_attrs(struct device *dev, struct sg_table *sgt,
321 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
322 		unsigned long attrs)
323 {
324 	const struct dma_map_ops *ops = get_dma_ops(dev);
325 
326 	if (dma_alloc_direct(dev, ops))
327 		return dma_direct_get_sgtable(dev, sgt, cpu_addr, dma_addr,
328 				size, attrs);
329 	if (!ops->get_sgtable)
330 		return -ENXIO;
331 	return ops->get_sgtable(dev, sgt, cpu_addr, dma_addr, size, attrs);
332 }
333 EXPORT_SYMBOL(dma_get_sgtable_attrs);
334 
335 #ifdef CONFIG_MMU
336 /*
337  * Return the page attributes used for mapping dma_alloc_* memory, either in
338  * kernel space if remapping is needed, or to userspace through dma_mmap_*.
339  */
340 pgprot_t dma_pgprot(struct device *dev, pgprot_t prot, unsigned long attrs)
341 {
342 	if (force_dma_unencrypted(dev))
343 		prot = pgprot_decrypted(prot);
344 	if (dev_is_dma_coherent(dev) ||
345 	    (IS_ENABLED(CONFIG_DMA_NONCOHERENT_CACHE_SYNC) &&
346              (attrs & DMA_ATTR_NON_CONSISTENT)))
347 		return prot;
348 #ifdef CONFIG_ARCH_HAS_DMA_WRITE_COMBINE
349 	if (attrs & DMA_ATTR_WRITE_COMBINE)
350 		return pgprot_writecombine(prot);
351 #endif
352 	return pgprot_dmacoherent(prot);
353 }
354 #endif /* CONFIG_MMU */
355 
356 /**
357  * dma_can_mmap - check if a given device supports dma_mmap_*
358  * @dev: device to check
359  *
360  * Returns %true if @dev supports dma_mmap_coherent() and dma_mmap_attrs() to
361  * map DMA allocations to userspace.
362  */
363 bool dma_can_mmap(struct device *dev)
364 {
365 	const struct dma_map_ops *ops = get_dma_ops(dev);
366 
367 	if (dma_alloc_direct(dev, ops))
368 		return dma_direct_can_mmap(dev);
369 	return ops->mmap != NULL;
370 }
371 EXPORT_SYMBOL_GPL(dma_can_mmap);
372 
373 /**
374  * dma_mmap_attrs - map a coherent DMA allocation into user space
375  * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
376  * @vma: vm_area_struct describing requested user mapping
377  * @cpu_addr: kernel CPU-view address returned from dma_alloc_attrs
378  * @dma_addr: device-view address returned from dma_alloc_attrs
379  * @size: size of memory originally requested in dma_alloc_attrs
380  * @attrs: attributes of mapping properties requested in dma_alloc_attrs
381  *
382  * Map a coherent DMA buffer previously allocated by dma_alloc_attrs into user
383  * space.  The coherent DMA buffer must not be freed by the driver until the
384  * user space mapping has been released.
385  */
386 int dma_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
387 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
388 		unsigned long attrs)
389 {
390 	const struct dma_map_ops *ops = get_dma_ops(dev);
391 
392 	if (dma_alloc_direct(dev, ops))
393 		return dma_direct_mmap(dev, vma, cpu_addr, dma_addr, size,
394 				attrs);
395 	if (!ops->mmap)
396 		return -ENXIO;
397 	return ops->mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
398 }
399 EXPORT_SYMBOL(dma_mmap_attrs);
400 
401 u64 dma_get_required_mask(struct device *dev)
402 {
403 	const struct dma_map_ops *ops = get_dma_ops(dev);
404 
405 	if (dma_alloc_direct(dev, ops))
406 		return dma_direct_get_required_mask(dev);
407 	if (ops->get_required_mask)
408 		return ops->get_required_mask(dev);
409 
410 	/*
411 	 * We require every DMA ops implementation to at least support a 32-bit
412 	 * DMA mask (and use bounce buffering if that isn't supported in
413 	 * hardware).  As the direct mapping code has its own routine to
414 	 * actually report an optimal mask we default to 32-bit here as that
415 	 * is the right thing for most IOMMUs, and at least not actively
416 	 * harmful in general.
417 	 */
418 	return DMA_BIT_MASK(32);
419 }
420 EXPORT_SYMBOL_GPL(dma_get_required_mask);
421 
422 void *dma_alloc_attrs(struct device *dev, size_t size, dma_addr_t *dma_handle,
423 		gfp_t flag, unsigned long attrs)
424 {
425 	const struct dma_map_ops *ops = get_dma_ops(dev);
426 	void *cpu_addr;
427 
428 	WARN_ON_ONCE(!dev->coherent_dma_mask);
429 
430 	if (dma_alloc_from_dev_coherent(dev, size, dma_handle, &cpu_addr))
431 		return cpu_addr;
432 
433 	/* let the implementation decide on the zone to allocate from: */
434 	flag &= ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM);
435 
436 	if (dma_alloc_direct(dev, ops))
437 		cpu_addr = dma_direct_alloc(dev, size, dma_handle, flag, attrs);
438 	else if (ops->alloc)
439 		cpu_addr = ops->alloc(dev, size, dma_handle, flag, attrs);
440 	else
441 		return NULL;
442 
443 	debug_dma_alloc_coherent(dev, size, *dma_handle, cpu_addr);
444 	return cpu_addr;
445 }
446 EXPORT_SYMBOL(dma_alloc_attrs);
447 
448 void dma_free_attrs(struct device *dev, size_t size, void *cpu_addr,
449 		dma_addr_t dma_handle, unsigned long attrs)
450 {
451 	const struct dma_map_ops *ops = get_dma_ops(dev);
452 
453 	if (dma_release_from_dev_coherent(dev, get_order(size), cpu_addr))
454 		return;
455 	/*
456 	 * On non-coherent platforms which implement DMA-coherent buffers via
457 	 * non-cacheable remaps, ops->free() may call vunmap(). Thus getting
458 	 * this far in IRQ context is a) at risk of a BUG_ON() or trying to
459 	 * sleep on some machines, and b) an indication that the driver is
460 	 * probably misusing the coherent API anyway.
461 	 */
462 	WARN_ON(irqs_disabled());
463 
464 	if (!cpu_addr)
465 		return;
466 
467 	debug_dma_free_coherent(dev, size, cpu_addr, dma_handle);
468 	if (dma_alloc_direct(dev, ops))
469 		dma_direct_free(dev, size, cpu_addr, dma_handle, attrs);
470 	else if (ops->free)
471 		ops->free(dev, size, cpu_addr, dma_handle, attrs);
472 }
473 EXPORT_SYMBOL(dma_free_attrs);
474 
475 int dma_supported(struct device *dev, u64 mask)
476 {
477 	const struct dma_map_ops *ops = get_dma_ops(dev);
478 
479 	/*
480 	 * ->dma_supported sets the bypass flag, so we must always call
481 	 * into the method here unless the device is truly direct mapped.
482 	 */
483 	if (!ops)
484 		return dma_direct_supported(dev, mask);
485 	if (!ops->dma_supported)
486 		return 1;
487 	return ops->dma_supported(dev, mask);
488 }
489 EXPORT_SYMBOL(dma_supported);
490 
491 #ifdef CONFIG_ARCH_HAS_DMA_SET_MASK
492 void arch_dma_set_mask(struct device *dev, u64 mask);
493 #else
494 #define arch_dma_set_mask(dev, mask)	do { } while (0)
495 #endif
496 
497 int dma_set_mask(struct device *dev, u64 mask)
498 {
499 	/*
500 	 * Truncate the mask to the actually supported dma_addr_t width to
501 	 * avoid generating unsupportable addresses.
502 	 */
503 	mask = (dma_addr_t)mask;
504 
505 	if (!dev->dma_mask || !dma_supported(dev, mask))
506 		return -EIO;
507 
508 	arch_dma_set_mask(dev, mask);
509 	*dev->dma_mask = mask;
510 	return 0;
511 }
512 EXPORT_SYMBOL(dma_set_mask);
513 
514 #ifndef CONFIG_ARCH_HAS_DMA_SET_COHERENT_MASK
515 int dma_set_coherent_mask(struct device *dev, u64 mask)
516 {
517 	/*
518 	 * Truncate the mask to the actually supported dma_addr_t width to
519 	 * avoid generating unsupportable addresses.
520 	 */
521 	mask = (dma_addr_t)mask;
522 
523 	if (!dma_supported(dev, mask))
524 		return -EIO;
525 
526 	dev->coherent_dma_mask = mask;
527 	return 0;
528 }
529 EXPORT_SYMBOL(dma_set_coherent_mask);
530 #endif
531 
532 void dma_cache_sync(struct device *dev, void *vaddr, size_t size,
533 		enum dma_data_direction dir)
534 {
535 	const struct dma_map_ops *ops = get_dma_ops(dev);
536 
537 	BUG_ON(!valid_dma_direction(dir));
538 
539 	if (dma_alloc_direct(dev, ops))
540 		arch_dma_cache_sync(dev, vaddr, size, dir);
541 	else if (ops->cache_sync)
542 		ops->cache_sync(dev, vaddr, size, dir);
543 }
544 EXPORT_SYMBOL(dma_cache_sync);
545 
546 size_t dma_max_mapping_size(struct device *dev)
547 {
548 	const struct dma_map_ops *ops = get_dma_ops(dev);
549 	size_t size = SIZE_MAX;
550 
551 	if (dma_map_direct(dev, ops))
552 		size = dma_direct_max_mapping_size(dev);
553 	else if (ops && ops->max_mapping_size)
554 		size = ops->max_mapping_size(dev);
555 
556 	return size;
557 }
558 EXPORT_SYMBOL_GPL(dma_max_mapping_size);
559 
560 bool dma_need_sync(struct device *dev, dma_addr_t dma_addr)
561 {
562 	const struct dma_map_ops *ops = get_dma_ops(dev);
563 
564 	if (dma_map_direct(dev, ops))
565 		return dma_direct_need_sync(dev, dma_addr);
566 	return ops->sync_single_for_cpu || ops->sync_single_for_device;
567 }
568 EXPORT_SYMBOL_GPL(dma_need_sync);
569 
570 unsigned long dma_get_merge_boundary(struct device *dev)
571 {
572 	const struct dma_map_ops *ops = get_dma_ops(dev);
573 
574 	if (!ops || !ops->get_merge_boundary)
575 		return 0;	/* can't merge */
576 
577 	return ops->get_merge_boundary(dev);
578 }
579 EXPORT_SYMBOL_GPL(dma_get_merge_boundary);
580