xref: /linux/arch/arm64/mm/dma-mapping.c (revision d69dece5f5b6bc7a5e39d2b6136ddc69469331fe)
1 /*
2  * SWIOTLB-based DMA API implementation
3  *
4  * Copyright (C) 2012 ARM Ltd.
5  * Author: Catalin Marinas <catalin.marinas@arm.com>
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License version 2 as
9  * published by the Free Software Foundation.
10  *
11  * This program is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
18  */
19 
20 #include <linux/gfp.h>
21 #include <linux/acpi.h>
22 #include <linux/bootmem.h>
23 #include <linux/cache.h>
24 #include <linux/export.h>
25 #include <linux/slab.h>
26 #include <linux/genalloc.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/dma-contiguous.h>
29 #include <linux/vmalloc.h>
30 #include <linux/swiotlb.h>
31 
32 #include <asm/cacheflush.h>
33 
34 static int swiotlb __ro_after_init;
35 
36 static pgprot_t __get_dma_pgprot(unsigned long attrs, pgprot_t prot,
37 				 bool coherent)
38 {
39 	if (!coherent || (attrs & DMA_ATTR_WRITE_COMBINE))
40 		return pgprot_writecombine(prot);
41 	return prot;
42 }
43 
44 static struct gen_pool *atomic_pool;
45 
46 #define DEFAULT_DMA_COHERENT_POOL_SIZE  SZ_256K
47 static size_t atomic_pool_size __initdata = DEFAULT_DMA_COHERENT_POOL_SIZE;
48 
49 static int __init early_coherent_pool(char *p)
50 {
51 	atomic_pool_size = memparse(p, &p);
52 	return 0;
53 }
54 early_param("coherent_pool", early_coherent_pool);
55 
56 static void *__alloc_from_pool(size_t size, struct page **ret_page, gfp_t flags)
57 {
58 	unsigned long val;
59 	void *ptr = NULL;
60 
61 	if (!atomic_pool) {
62 		WARN(1, "coherent pool not initialised!\n");
63 		return NULL;
64 	}
65 
66 	val = gen_pool_alloc(atomic_pool, size);
67 	if (val) {
68 		phys_addr_t phys = gen_pool_virt_to_phys(atomic_pool, val);
69 
70 		*ret_page = phys_to_page(phys);
71 		ptr = (void *)val;
72 		memset(ptr, 0, size);
73 	}
74 
75 	return ptr;
76 }
77 
78 static bool __in_atomic_pool(void *start, size_t size)
79 {
80 	return addr_in_gen_pool(atomic_pool, (unsigned long)start, size);
81 }
82 
83 static int __free_from_pool(void *start, size_t size)
84 {
85 	if (!__in_atomic_pool(start, size))
86 		return 0;
87 
88 	gen_pool_free(atomic_pool, (unsigned long)start, size);
89 
90 	return 1;
91 }
92 
93 static void *__dma_alloc_coherent(struct device *dev, size_t size,
94 				  dma_addr_t *dma_handle, gfp_t flags,
95 				  unsigned long attrs)
96 {
97 	if (dev == NULL) {
98 		WARN_ONCE(1, "Use an actual device structure for DMA allocation\n");
99 		return NULL;
100 	}
101 
102 	if (IS_ENABLED(CONFIG_ZONE_DMA) &&
103 	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
104 		flags |= GFP_DMA;
105 	if (dev_get_cma_area(dev) && gfpflags_allow_blocking(flags)) {
106 		struct page *page;
107 		void *addr;
108 
109 		page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
110 							get_order(size));
111 		if (!page)
112 			return NULL;
113 
114 		*dma_handle = phys_to_dma(dev, page_to_phys(page));
115 		addr = page_address(page);
116 		memset(addr, 0, size);
117 		return addr;
118 	} else {
119 		return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
120 	}
121 }
122 
123 static void __dma_free_coherent(struct device *dev, size_t size,
124 				void *vaddr, dma_addr_t dma_handle,
125 				unsigned long attrs)
126 {
127 	bool freed;
128 	phys_addr_t paddr = dma_to_phys(dev, dma_handle);
129 
130 	if (dev == NULL) {
131 		WARN_ONCE(1, "Use an actual device structure for DMA allocation\n");
132 		return;
133 	}
134 
135 	freed = dma_release_from_contiguous(dev,
136 					phys_to_page(paddr),
137 					size >> PAGE_SHIFT);
138 	if (!freed)
139 		swiotlb_free_coherent(dev, size, vaddr, dma_handle);
140 }
141 
142 static void *__dma_alloc(struct device *dev, size_t size,
143 			 dma_addr_t *dma_handle, gfp_t flags,
144 			 unsigned long attrs)
145 {
146 	struct page *page;
147 	void *ptr, *coherent_ptr;
148 	bool coherent = is_device_dma_coherent(dev);
149 	pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, false);
150 
151 	size = PAGE_ALIGN(size);
152 
153 	if (!coherent && !gfpflags_allow_blocking(flags)) {
154 		struct page *page = NULL;
155 		void *addr = __alloc_from_pool(size, &page, flags);
156 
157 		if (addr)
158 			*dma_handle = phys_to_dma(dev, page_to_phys(page));
159 
160 		return addr;
161 	}
162 
163 	ptr = __dma_alloc_coherent(dev, size, dma_handle, flags, attrs);
164 	if (!ptr)
165 		goto no_mem;
166 
167 	/* no need for non-cacheable mapping if coherent */
168 	if (coherent)
169 		return ptr;
170 
171 	/* remove any dirty cache lines on the kernel alias */
172 	__dma_flush_area(ptr, size);
173 
174 	/* create a coherent mapping */
175 	page = virt_to_page(ptr);
176 	coherent_ptr = dma_common_contiguous_remap(page, size, VM_USERMAP,
177 						   prot, NULL);
178 	if (!coherent_ptr)
179 		goto no_map;
180 
181 	return coherent_ptr;
182 
183 no_map:
184 	__dma_free_coherent(dev, size, ptr, *dma_handle, attrs);
185 no_mem:
186 	*dma_handle = DMA_ERROR_CODE;
187 	return NULL;
188 }
189 
190 static void __dma_free(struct device *dev, size_t size,
191 		       void *vaddr, dma_addr_t dma_handle,
192 		       unsigned long attrs)
193 {
194 	void *swiotlb_addr = phys_to_virt(dma_to_phys(dev, dma_handle));
195 
196 	size = PAGE_ALIGN(size);
197 
198 	if (!is_device_dma_coherent(dev)) {
199 		if (__free_from_pool(vaddr, size))
200 			return;
201 		vunmap(vaddr);
202 	}
203 	__dma_free_coherent(dev, size, swiotlb_addr, dma_handle, attrs);
204 }
205 
206 static dma_addr_t __swiotlb_map_page(struct device *dev, struct page *page,
207 				     unsigned long offset, size_t size,
208 				     enum dma_data_direction dir,
209 				     unsigned long attrs)
210 {
211 	dma_addr_t dev_addr;
212 
213 	dev_addr = swiotlb_map_page(dev, page, offset, size, dir, attrs);
214 	if (!is_device_dma_coherent(dev))
215 		__dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
216 
217 	return dev_addr;
218 }
219 
220 
221 static void __swiotlb_unmap_page(struct device *dev, dma_addr_t dev_addr,
222 				 size_t size, enum dma_data_direction dir,
223 				 unsigned long attrs)
224 {
225 	if (!is_device_dma_coherent(dev))
226 		__dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
227 	swiotlb_unmap_page(dev, dev_addr, size, dir, attrs);
228 }
229 
230 static int __swiotlb_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
231 				  int nelems, enum dma_data_direction dir,
232 				  unsigned long attrs)
233 {
234 	struct scatterlist *sg;
235 	int i, ret;
236 
237 	ret = swiotlb_map_sg_attrs(dev, sgl, nelems, dir, attrs);
238 	if (!is_device_dma_coherent(dev))
239 		for_each_sg(sgl, sg, ret, i)
240 			__dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
241 				       sg->length, dir);
242 
243 	return ret;
244 }
245 
246 static void __swiotlb_unmap_sg_attrs(struct device *dev,
247 				     struct scatterlist *sgl, int nelems,
248 				     enum dma_data_direction dir,
249 				     unsigned long attrs)
250 {
251 	struct scatterlist *sg;
252 	int i;
253 
254 	if (!is_device_dma_coherent(dev))
255 		for_each_sg(sgl, sg, nelems, i)
256 			__dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
257 					 sg->length, dir);
258 	swiotlb_unmap_sg_attrs(dev, sgl, nelems, dir, attrs);
259 }
260 
261 static void __swiotlb_sync_single_for_cpu(struct device *dev,
262 					  dma_addr_t dev_addr, size_t size,
263 					  enum dma_data_direction dir)
264 {
265 	if (!is_device_dma_coherent(dev))
266 		__dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
267 	swiotlb_sync_single_for_cpu(dev, dev_addr, size, dir);
268 }
269 
270 static void __swiotlb_sync_single_for_device(struct device *dev,
271 					     dma_addr_t dev_addr, size_t size,
272 					     enum dma_data_direction dir)
273 {
274 	swiotlb_sync_single_for_device(dev, dev_addr, size, dir);
275 	if (!is_device_dma_coherent(dev))
276 		__dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
277 }
278 
279 static void __swiotlb_sync_sg_for_cpu(struct device *dev,
280 				      struct scatterlist *sgl, int nelems,
281 				      enum dma_data_direction dir)
282 {
283 	struct scatterlist *sg;
284 	int i;
285 
286 	if (!is_device_dma_coherent(dev))
287 		for_each_sg(sgl, sg, nelems, i)
288 			__dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
289 					 sg->length, dir);
290 	swiotlb_sync_sg_for_cpu(dev, sgl, nelems, dir);
291 }
292 
293 static void __swiotlb_sync_sg_for_device(struct device *dev,
294 					 struct scatterlist *sgl, int nelems,
295 					 enum dma_data_direction dir)
296 {
297 	struct scatterlist *sg;
298 	int i;
299 
300 	swiotlb_sync_sg_for_device(dev, sgl, nelems, dir);
301 	if (!is_device_dma_coherent(dev))
302 		for_each_sg(sgl, sg, nelems, i)
303 			__dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
304 				       sg->length, dir);
305 }
306 
307 static int __swiotlb_mmap(struct device *dev,
308 			  struct vm_area_struct *vma,
309 			  void *cpu_addr, dma_addr_t dma_addr, size_t size,
310 			  unsigned long attrs)
311 {
312 	int ret = -ENXIO;
313 	unsigned long nr_vma_pages = (vma->vm_end - vma->vm_start) >>
314 					PAGE_SHIFT;
315 	unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
316 	unsigned long pfn = dma_to_phys(dev, dma_addr) >> PAGE_SHIFT;
317 	unsigned long off = vma->vm_pgoff;
318 
319 	vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot,
320 					     is_device_dma_coherent(dev));
321 
322 	if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
323 		return ret;
324 
325 	if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) {
326 		ret = remap_pfn_range(vma, vma->vm_start,
327 				      pfn + off,
328 				      vma->vm_end - vma->vm_start,
329 				      vma->vm_page_prot);
330 	}
331 
332 	return ret;
333 }
334 
335 static int __swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
336 				 void *cpu_addr, dma_addr_t handle, size_t size,
337 				 unsigned long attrs)
338 {
339 	int ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
340 
341 	if (!ret)
342 		sg_set_page(sgt->sgl, phys_to_page(dma_to_phys(dev, handle)),
343 			    PAGE_ALIGN(size), 0);
344 
345 	return ret;
346 }
347 
348 static int __swiotlb_dma_supported(struct device *hwdev, u64 mask)
349 {
350 	if (swiotlb)
351 		return swiotlb_dma_supported(hwdev, mask);
352 	return 1;
353 }
354 
355 static struct dma_map_ops swiotlb_dma_ops = {
356 	.alloc = __dma_alloc,
357 	.free = __dma_free,
358 	.mmap = __swiotlb_mmap,
359 	.get_sgtable = __swiotlb_get_sgtable,
360 	.map_page = __swiotlb_map_page,
361 	.unmap_page = __swiotlb_unmap_page,
362 	.map_sg = __swiotlb_map_sg_attrs,
363 	.unmap_sg = __swiotlb_unmap_sg_attrs,
364 	.sync_single_for_cpu = __swiotlb_sync_single_for_cpu,
365 	.sync_single_for_device = __swiotlb_sync_single_for_device,
366 	.sync_sg_for_cpu = __swiotlb_sync_sg_for_cpu,
367 	.sync_sg_for_device = __swiotlb_sync_sg_for_device,
368 	.dma_supported = __swiotlb_dma_supported,
369 	.mapping_error = swiotlb_dma_mapping_error,
370 };
371 
372 static int __init atomic_pool_init(void)
373 {
374 	pgprot_t prot = __pgprot(PROT_NORMAL_NC);
375 	unsigned long nr_pages = atomic_pool_size >> PAGE_SHIFT;
376 	struct page *page;
377 	void *addr;
378 	unsigned int pool_size_order = get_order(atomic_pool_size);
379 
380 	if (dev_get_cma_area(NULL))
381 		page = dma_alloc_from_contiguous(NULL, nr_pages,
382 							pool_size_order);
383 	else
384 		page = alloc_pages(GFP_DMA, pool_size_order);
385 
386 	if (page) {
387 		int ret;
388 		void *page_addr = page_address(page);
389 
390 		memset(page_addr, 0, atomic_pool_size);
391 		__dma_flush_area(page_addr, atomic_pool_size);
392 
393 		atomic_pool = gen_pool_create(PAGE_SHIFT, -1);
394 		if (!atomic_pool)
395 			goto free_page;
396 
397 		addr = dma_common_contiguous_remap(page, atomic_pool_size,
398 					VM_USERMAP, prot, atomic_pool_init);
399 
400 		if (!addr)
401 			goto destroy_genpool;
402 
403 		ret = gen_pool_add_virt(atomic_pool, (unsigned long)addr,
404 					page_to_phys(page),
405 					atomic_pool_size, -1);
406 		if (ret)
407 			goto remove_mapping;
408 
409 		gen_pool_set_algo(atomic_pool,
410 				  gen_pool_first_fit_order_align,
411 				  (void *)PAGE_SHIFT);
412 
413 		pr_info("DMA: preallocated %zu KiB pool for atomic allocations\n",
414 			atomic_pool_size / 1024);
415 		return 0;
416 	}
417 	goto out;
418 
419 remove_mapping:
420 	dma_common_free_remap(addr, atomic_pool_size, VM_USERMAP);
421 destroy_genpool:
422 	gen_pool_destroy(atomic_pool);
423 	atomic_pool = NULL;
424 free_page:
425 	if (!dma_release_from_contiguous(NULL, page, nr_pages))
426 		__free_pages(page, pool_size_order);
427 out:
428 	pr_err("DMA: failed to allocate %zu KiB pool for atomic coherent allocation\n",
429 		atomic_pool_size / 1024);
430 	return -ENOMEM;
431 }
432 
433 /********************************************
434  * The following APIs are for dummy DMA ops *
435  ********************************************/
436 
437 static void *__dummy_alloc(struct device *dev, size_t size,
438 			   dma_addr_t *dma_handle, gfp_t flags,
439 			   unsigned long attrs)
440 {
441 	return NULL;
442 }
443 
444 static void __dummy_free(struct device *dev, size_t size,
445 			 void *vaddr, dma_addr_t dma_handle,
446 			 unsigned long attrs)
447 {
448 }
449 
450 static int __dummy_mmap(struct device *dev,
451 			struct vm_area_struct *vma,
452 			void *cpu_addr, dma_addr_t dma_addr, size_t size,
453 			unsigned long attrs)
454 {
455 	return -ENXIO;
456 }
457 
458 static dma_addr_t __dummy_map_page(struct device *dev, struct page *page,
459 				   unsigned long offset, size_t size,
460 				   enum dma_data_direction dir,
461 				   unsigned long attrs)
462 {
463 	return DMA_ERROR_CODE;
464 }
465 
466 static void __dummy_unmap_page(struct device *dev, dma_addr_t dev_addr,
467 			       size_t size, enum dma_data_direction dir,
468 			       unsigned long attrs)
469 {
470 }
471 
472 static int __dummy_map_sg(struct device *dev, struct scatterlist *sgl,
473 			  int nelems, enum dma_data_direction dir,
474 			  unsigned long attrs)
475 {
476 	return 0;
477 }
478 
479 static void __dummy_unmap_sg(struct device *dev,
480 			     struct scatterlist *sgl, int nelems,
481 			     enum dma_data_direction dir,
482 			     unsigned long attrs)
483 {
484 }
485 
486 static void __dummy_sync_single(struct device *dev,
487 				dma_addr_t dev_addr, size_t size,
488 				enum dma_data_direction dir)
489 {
490 }
491 
492 static void __dummy_sync_sg(struct device *dev,
493 			    struct scatterlist *sgl, int nelems,
494 			    enum dma_data_direction dir)
495 {
496 }
497 
498 static int __dummy_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
499 {
500 	return 1;
501 }
502 
503 static int __dummy_dma_supported(struct device *hwdev, u64 mask)
504 {
505 	return 0;
506 }
507 
508 struct dma_map_ops dummy_dma_ops = {
509 	.alloc                  = __dummy_alloc,
510 	.free                   = __dummy_free,
511 	.mmap                   = __dummy_mmap,
512 	.map_page               = __dummy_map_page,
513 	.unmap_page             = __dummy_unmap_page,
514 	.map_sg                 = __dummy_map_sg,
515 	.unmap_sg               = __dummy_unmap_sg,
516 	.sync_single_for_cpu    = __dummy_sync_single,
517 	.sync_single_for_device = __dummy_sync_single,
518 	.sync_sg_for_cpu        = __dummy_sync_sg,
519 	.sync_sg_for_device     = __dummy_sync_sg,
520 	.mapping_error          = __dummy_mapping_error,
521 	.dma_supported          = __dummy_dma_supported,
522 };
523 EXPORT_SYMBOL(dummy_dma_ops);
524 
525 static int __init arm64_dma_init(void)
526 {
527 	if (swiotlb_force == SWIOTLB_FORCE ||
528 	    max_pfn > (arm64_dma_phys_limit >> PAGE_SHIFT))
529 		swiotlb = 1;
530 
531 	return atomic_pool_init();
532 }
533 arch_initcall(arm64_dma_init);
534 
535 #define PREALLOC_DMA_DEBUG_ENTRIES	4096
536 
537 static int __init dma_debug_do_init(void)
538 {
539 	dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
540 	return 0;
541 }
542 fs_initcall(dma_debug_do_init);
543 
544 
545 #ifdef CONFIG_IOMMU_DMA
546 #include <linux/dma-iommu.h>
547 #include <linux/platform_device.h>
548 #include <linux/amba/bus.h>
549 
550 /* Thankfully, all cache ops are by VA so we can ignore phys here */
551 static void flush_page(struct device *dev, const void *virt, phys_addr_t phys)
552 {
553 	__dma_flush_area(virt, PAGE_SIZE);
554 }
555 
556 static void *__iommu_alloc_attrs(struct device *dev, size_t size,
557 				 dma_addr_t *handle, gfp_t gfp,
558 				 unsigned long attrs)
559 {
560 	bool coherent = is_device_dma_coherent(dev);
561 	int ioprot = dma_direction_to_prot(DMA_BIDIRECTIONAL, coherent);
562 	size_t iosize = size;
563 	void *addr;
564 
565 	if (WARN(!dev, "cannot create IOMMU mapping for unknown device\n"))
566 		return NULL;
567 
568 	size = PAGE_ALIGN(size);
569 
570 	/*
571 	 * Some drivers rely on this, and we probably don't want the
572 	 * possibility of stale kernel data being read by devices anyway.
573 	 */
574 	gfp |= __GFP_ZERO;
575 
576 	if (gfpflags_allow_blocking(gfp)) {
577 		struct page **pages;
578 		pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, coherent);
579 
580 		pages = iommu_dma_alloc(dev, iosize, gfp, attrs, ioprot,
581 					handle, flush_page);
582 		if (!pages)
583 			return NULL;
584 
585 		addr = dma_common_pages_remap(pages, size, VM_USERMAP, prot,
586 					      __builtin_return_address(0));
587 		if (!addr)
588 			iommu_dma_free(dev, pages, iosize, handle);
589 	} else {
590 		struct page *page;
591 		/*
592 		 * In atomic context we can't remap anything, so we'll only
593 		 * get the virtually contiguous buffer we need by way of a
594 		 * physically contiguous allocation.
595 		 */
596 		if (coherent) {
597 			page = alloc_pages(gfp, get_order(size));
598 			addr = page ? page_address(page) : NULL;
599 		} else {
600 			addr = __alloc_from_pool(size, &page, gfp);
601 		}
602 		if (!addr)
603 			return NULL;
604 
605 		*handle = iommu_dma_map_page(dev, page, 0, iosize, ioprot);
606 		if (iommu_dma_mapping_error(dev, *handle)) {
607 			if (coherent)
608 				__free_pages(page, get_order(size));
609 			else
610 				__free_from_pool(addr, size);
611 			addr = NULL;
612 		}
613 	}
614 	return addr;
615 }
616 
617 static void __iommu_free_attrs(struct device *dev, size_t size, void *cpu_addr,
618 			       dma_addr_t handle, unsigned long attrs)
619 {
620 	size_t iosize = size;
621 
622 	size = PAGE_ALIGN(size);
623 	/*
624 	 * @cpu_addr will be one of 3 things depending on how it was allocated:
625 	 * - A remapped array of pages from iommu_dma_alloc(), for all
626 	 *   non-atomic allocations.
627 	 * - A non-cacheable alias from the atomic pool, for atomic
628 	 *   allocations by non-coherent devices.
629 	 * - A normal lowmem address, for atomic allocations by
630 	 *   coherent devices.
631 	 * Hence how dodgy the below logic looks...
632 	 */
633 	if (__in_atomic_pool(cpu_addr, size)) {
634 		iommu_dma_unmap_page(dev, handle, iosize, 0, 0);
635 		__free_from_pool(cpu_addr, size);
636 	} else if (is_vmalloc_addr(cpu_addr)){
637 		struct vm_struct *area = find_vm_area(cpu_addr);
638 
639 		if (WARN_ON(!area || !area->pages))
640 			return;
641 		iommu_dma_free(dev, area->pages, iosize, &handle);
642 		dma_common_free_remap(cpu_addr, size, VM_USERMAP);
643 	} else {
644 		iommu_dma_unmap_page(dev, handle, iosize, 0, 0);
645 		__free_pages(virt_to_page(cpu_addr), get_order(size));
646 	}
647 }
648 
649 static int __iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
650 			      void *cpu_addr, dma_addr_t dma_addr, size_t size,
651 			      unsigned long attrs)
652 {
653 	struct vm_struct *area;
654 	int ret;
655 
656 	vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot,
657 					     is_device_dma_coherent(dev));
658 
659 	if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
660 		return ret;
661 
662 	area = find_vm_area(cpu_addr);
663 	if (WARN_ON(!area || !area->pages))
664 		return -ENXIO;
665 
666 	return iommu_dma_mmap(area->pages, size, vma);
667 }
668 
669 static int __iommu_get_sgtable(struct device *dev, struct sg_table *sgt,
670 			       void *cpu_addr, dma_addr_t dma_addr,
671 			       size_t size, unsigned long attrs)
672 {
673 	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
674 	struct vm_struct *area = find_vm_area(cpu_addr);
675 
676 	if (WARN_ON(!area || !area->pages))
677 		return -ENXIO;
678 
679 	return sg_alloc_table_from_pages(sgt, area->pages, count, 0, size,
680 					 GFP_KERNEL);
681 }
682 
683 static void __iommu_sync_single_for_cpu(struct device *dev,
684 					dma_addr_t dev_addr, size_t size,
685 					enum dma_data_direction dir)
686 {
687 	phys_addr_t phys;
688 
689 	if (is_device_dma_coherent(dev))
690 		return;
691 
692 	phys = iommu_iova_to_phys(iommu_get_domain_for_dev(dev), dev_addr);
693 	__dma_unmap_area(phys_to_virt(phys), size, dir);
694 }
695 
696 static void __iommu_sync_single_for_device(struct device *dev,
697 					   dma_addr_t dev_addr, size_t size,
698 					   enum dma_data_direction dir)
699 {
700 	phys_addr_t phys;
701 
702 	if (is_device_dma_coherent(dev))
703 		return;
704 
705 	phys = iommu_iova_to_phys(iommu_get_domain_for_dev(dev), dev_addr);
706 	__dma_map_area(phys_to_virt(phys), size, dir);
707 }
708 
709 static dma_addr_t __iommu_map_page(struct device *dev, struct page *page,
710 				   unsigned long offset, size_t size,
711 				   enum dma_data_direction dir,
712 				   unsigned long attrs)
713 {
714 	bool coherent = is_device_dma_coherent(dev);
715 	int prot = dma_direction_to_prot(dir, coherent);
716 	dma_addr_t dev_addr = iommu_dma_map_page(dev, page, offset, size, prot);
717 
718 	if (!iommu_dma_mapping_error(dev, dev_addr) &&
719 	    (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
720 		__iommu_sync_single_for_device(dev, dev_addr, size, dir);
721 
722 	return dev_addr;
723 }
724 
725 static void __iommu_unmap_page(struct device *dev, dma_addr_t dev_addr,
726 			       size_t size, enum dma_data_direction dir,
727 			       unsigned long attrs)
728 {
729 	if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
730 		__iommu_sync_single_for_cpu(dev, dev_addr, size, dir);
731 
732 	iommu_dma_unmap_page(dev, dev_addr, size, dir, attrs);
733 }
734 
735 static void __iommu_sync_sg_for_cpu(struct device *dev,
736 				    struct scatterlist *sgl, int nelems,
737 				    enum dma_data_direction dir)
738 {
739 	struct scatterlist *sg;
740 	int i;
741 
742 	if (is_device_dma_coherent(dev))
743 		return;
744 
745 	for_each_sg(sgl, sg, nelems, i)
746 		__dma_unmap_area(sg_virt(sg), sg->length, dir);
747 }
748 
749 static void __iommu_sync_sg_for_device(struct device *dev,
750 				       struct scatterlist *sgl, int nelems,
751 				       enum dma_data_direction dir)
752 {
753 	struct scatterlist *sg;
754 	int i;
755 
756 	if (is_device_dma_coherent(dev))
757 		return;
758 
759 	for_each_sg(sgl, sg, nelems, i)
760 		__dma_map_area(sg_virt(sg), sg->length, dir);
761 }
762 
763 static int __iommu_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
764 				int nelems, enum dma_data_direction dir,
765 				unsigned long attrs)
766 {
767 	bool coherent = is_device_dma_coherent(dev);
768 
769 	if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
770 		__iommu_sync_sg_for_device(dev, sgl, nelems, dir);
771 
772 	return iommu_dma_map_sg(dev, sgl, nelems,
773 			dma_direction_to_prot(dir, coherent));
774 }
775 
776 static void __iommu_unmap_sg_attrs(struct device *dev,
777 				   struct scatterlist *sgl, int nelems,
778 				   enum dma_data_direction dir,
779 				   unsigned long attrs)
780 {
781 	if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
782 		__iommu_sync_sg_for_cpu(dev, sgl, nelems, dir);
783 
784 	iommu_dma_unmap_sg(dev, sgl, nelems, dir, attrs);
785 }
786 
787 static struct dma_map_ops iommu_dma_ops = {
788 	.alloc = __iommu_alloc_attrs,
789 	.free = __iommu_free_attrs,
790 	.mmap = __iommu_mmap_attrs,
791 	.get_sgtable = __iommu_get_sgtable,
792 	.map_page = __iommu_map_page,
793 	.unmap_page = __iommu_unmap_page,
794 	.map_sg = __iommu_map_sg_attrs,
795 	.unmap_sg = __iommu_unmap_sg_attrs,
796 	.sync_single_for_cpu = __iommu_sync_single_for_cpu,
797 	.sync_single_for_device = __iommu_sync_single_for_device,
798 	.sync_sg_for_cpu = __iommu_sync_sg_for_cpu,
799 	.sync_sg_for_device = __iommu_sync_sg_for_device,
800 	.map_resource = iommu_dma_map_resource,
801 	.unmap_resource = iommu_dma_unmap_resource,
802 	.dma_supported = iommu_dma_supported,
803 	.mapping_error = iommu_dma_mapping_error,
804 };
805 
806 /*
807  * TODO: Right now __iommu_setup_dma_ops() gets called too early to do
808  * everything it needs to - the device is only partially created and the
809  * IOMMU driver hasn't seen it yet, so it can't have a group. Thus we
810  * need this delayed attachment dance. Once IOMMU probe ordering is sorted
811  * to move the arch_setup_dma_ops() call later, all the notifier bits below
812  * become unnecessary, and will go away.
813  */
814 struct iommu_dma_notifier_data {
815 	struct list_head list;
816 	struct device *dev;
817 	const struct iommu_ops *ops;
818 	u64 dma_base;
819 	u64 size;
820 };
821 static LIST_HEAD(iommu_dma_masters);
822 static DEFINE_MUTEX(iommu_dma_notifier_lock);
823 
824 static bool do_iommu_attach(struct device *dev, const struct iommu_ops *ops,
825 			   u64 dma_base, u64 size)
826 {
827 	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
828 
829 	/*
830 	 * If the IOMMU driver has the DMA domain support that we require,
831 	 * then the IOMMU core will have already configured a group for this
832 	 * device, and allocated the default domain for that group.
833 	 */
834 	if (!domain || iommu_dma_init_domain(domain, dma_base, size, dev)) {
835 		pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n",
836 			dev_name(dev));
837 		return false;
838 	}
839 
840 	dev->archdata.dma_ops = &iommu_dma_ops;
841 	return true;
842 }
843 
844 static void queue_iommu_attach(struct device *dev, const struct iommu_ops *ops,
845 			      u64 dma_base, u64 size)
846 {
847 	struct iommu_dma_notifier_data *iommudata;
848 
849 	iommudata = kzalloc(sizeof(*iommudata), GFP_KERNEL);
850 	if (!iommudata)
851 		return;
852 
853 	iommudata->dev = dev;
854 	iommudata->ops = ops;
855 	iommudata->dma_base = dma_base;
856 	iommudata->size = size;
857 
858 	mutex_lock(&iommu_dma_notifier_lock);
859 	list_add(&iommudata->list, &iommu_dma_masters);
860 	mutex_unlock(&iommu_dma_notifier_lock);
861 }
862 
863 static int __iommu_attach_notifier(struct notifier_block *nb,
864 				   unsigned long action, void *data)
865 {
866 	struct iommu_dma_notifier_data *master, *tmp;
867 
868 	if (action != BUS_NOTIFY_BIND_DRIVER)
869 		return 0;
870 
871 	mutex_lock(&iommu_dma_notifier_lock);
872 	list_for_each_entry_safe(master, tmp, &iommu_dma_masters, list) {
873 		if (data == master->dev && do_iommu_attach(master->dev,
874 				master->ops, master->dma_base, master->size)) {
875 			list_del(&master->list);
876 			kfree(master);
877 			break;
878 		}
879 	}
880 	mutex_unlock(&iommu_dma_notifier_lock);
881 	return 0;
882 }
883 
884 static int __init register_iommu_dma_ops_notifier(struct bus_type *bus)
885 {
886 	struct notifier_block *nb = kzalloc(sizeof(*nb), GFP_KERNEL);
887 	int ret;
888 
889 	if (!nb)
890 		return -ENOMEM;
891 
892 	nb->notifier_call = __iommu_attach_notifier;
893 
894 	ret = bus_register_notifier(bus, nb);
895 	if (ret) {
896 		pr_warn("Failed to register DMA domain notifier; IOMMU DMA ops unavailable on bus '%s'\n",
897 			bus->name);
898 		kfree(nb);
899 	}
900 	return ret;
901 }
902 
903 static int __init __iommu_dma_init(void)
904 {
905 	int ret;
906 
907 	ret = iommu_dma_init();
908 	if (!ret)
909 		ret = register_iommu_dma_ops_notifier(&platform_bus_type);
910 	if (!ret)
911 		ret = register_iommu_dma_ops_notifier(&amba_bustype);
912 #ifdef CONFIG_PCI
913 	if (!ret)
914 		ret = register_iommu_dma_ops_notifier(&pci_bus_type);
915 #endif
916 	return ret;
917 }
918 arch_initcall(__iommu_dma_init);
919 
920 static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
921 				  const struct iommu_ops *ops)
922 {
923 	struct iommu_group *group;
924 
925 	if (!ops)
926 		return;
927 	/*
928 	 * TODO: As a concession to the future, we're ready to handle being
929 	 * called both early and late (i.e. after bus_add_device). Once all
930 	 * the platform bus code is reworked to call us late and the notifier
931 	 * junk above goes away, move the body of do_iommu_attach here.
932 	 */
933 	group = iommu_group_get(dev);
934 	if (group) {
935 		do_iommu_attach(dev, ops, dma_base, size);
936 		iommu_group_put(group);
937 	} else {
938 		queue_iommu_attach(dev, ops, dma_base, size);
939 	}
940 }
941 
942 void arch_teardown_dma_ops(struct device *dev)
943 {
944 	dev->archdata.dma_ops = NULL;
945 }
946 
947 #else
948 
949 static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
950 				  const struct iommu_ops *iommu)
951 { }
952 
953 #endif  /* CONFIG_IOMMU_DMA */
954 
955 void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
956 			const struct iommu_ops *iommu, bool coherent)
957 {
958 	if (!dev->archdata.dma_ops)
959 		dev->archdata.dma_ops = &swiotlb_dma_ops;
960 
961 	dev->archdata.dma_coherent = coherent;
962 	__iommu_setup_dma_ops(dev, dma_base, size, iommu);
963 }
964