Commit | Line | Data |
---|---|---|
0db2e5d1 RM |
1 | /* |
2 | * A fairly generic DMA-API to IOMMU-API glue layer. | |
3 | * | |
4 | * Copyright (C) 2014-2015 ARM Ltd. | |
5 | * | |
6 | * based in part on arch/arm/mm/dma-mapping.c: | |
7 | * Copyright (C) 2000-2004 Russell King | |
8 | * | |
9 | * This program is free software; you can redistribute it and/or modify | |
10 | * it under the terms of the GNU General Public License version 2 as | |
11 | * published by the Free Software Foundation. | |
12 | * | |
13 | * This program is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
16 | * GNU General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU General Public License | |
19 | * along with this program. If not, see <http://www.gnu.org/licenses/>. | |
20 | */ | |
21 | ||
22 | #include <linux/device.h> | |
23 | #include <linux/dma-iommu.h> | |
5b11e9cd | 24 | #include <linux/gfp.h> |
0db2e5d1 RM |
25 | #include <linux/huge_mm.h> |
26 | #include <linux/iommu.h> | |
27 | #include <linux/iova.h> | |
28 | #include <linux/mm.h> | |
5b11e9cd RM |
29 | #include <linux/scatterlist.h> |
30 | #include <linux/vmalloc.h> | |
0db2e5d1 RM |
31 | |
32 | int iommu_dma_init(void) | |
33 | { | |
34 | return iova_cache_get(); | |
35 | } | |
36 | ||
37 | /** | |
38 | * iommu_get_dma_cookie - Acquire DMA-API resources for a domain | |
39 | * @domain: IOMMU domain to prepare for DMA-API usage | |
40 | * | |
41 | * IOMMU drivers should normally call this from their domain_alloc | |
42 | * callback when domain->type == IOMMU_DOMAIN_DMA. | |
43 | */ | |
44 | int iommu_get_dma_cookie(struct iommu_domain *domain) | |
45 | { | |
46 | struct iova_domain *iovad; | |
47 | ||
48 | if (domain->iova_cookie) | |
49 | return -EEXIST; | |
50 | ||
51 | iovad = kzalloc(sizeof(*iovad), GFP_KERNEL); | |
52 | domain->iova_cookie = iovad; | |
53 | ||
54 | return iovad ? 0 : -ENOMEM; | |
55 | } | |
56 | EXPORT_SYMBOL(iommu_get_dma_cookie); | |
57 | ||
58 | /** | |
59 | * iommu_put_dma_cookie - Release a domain's DMA mapping resources | |
60 | * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie() | |
61 | * | |
62 | * IOMMU drivers should normally call this from their domain_free callback. | |
63 | */ | |
64 | void iommu_put_dma_cookie(struct iommu_domain *domain) | |
65 | { | |
66 | struct iova_domain *iovad = domain->iova_cookie; | |
67 | ||
68 | if (!iovad) | |
69 | return; | |
70 | ||
71 | put_iova_domain(iovad); | |
72 | kfree(iovad); | |
73 | domain->iova_cookie = NULL; | |
74 | } | |
75 | EXPORT_SYMBOL(iommu_put_dma_cookie); | |
76 | ||
77 | /** | |
78 | * iommu_dma_init_domain - Initialise a DMA mapping domain | |
79 | * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie() | |
80 | * @base: IOVA at which the mappable address space starts | |
81 | * @size: Size of IOVA space | |
82 | * | |
83 | * @base and @size should be exact multiples of IOMMU page granularity to | |
84 | * avoid rounding surprises. If necessary, we reserve the page at address 0 | |
85 | * to ensure it is an invalid IOVA. It is safe to reinitialise a domain, but | |
86 | * any change which could make prior IOVAs invalid will fail. | |
87 | */ | |
88 | int iommu_dma_init_domain(struct iommu_domain *domain, dma_addr_t base, u64 size) | |
89 | { | |
90 | struct iova_domain *iovad = domain->iova_cookie; | |
91 | unsigned long order, base_pfn, end_pfn; | |
92 | ||
93 | if (!iovad) | |
94 | return -ENODEV; | |
95 | ||
96 | /* Use the smallest supported page size for IOVA granularity */ | |
97 | order = __ffs(domain->ops->pgsize_bitmap); | |
98 | base_pfn = max_t(unsigned long, 1, base >> order); | |
99 | end_pfn = (base + size - 1) >> order; | |
100 | ||
101 | /* Check the domain allows at least some access to the device... */ | |
102 | if (domain->geometry.force_aperture) { | |
103 | if (base > domain->geometry.aperture_end || | |
104 | base + size <= domain->geometry.aperture_start) { | |
105 | pr_warn("specified DMA range outside IOMMU capability\n"); | |
106 | return -EFAULT; | |
107 | } | |
108 | /* ...then finally give it a kicking to make sure it fits */ | |
109 | base_pfn = max_t(unsigned long, base_pfn, | |
110 | domain->geometry.aperture_start >> order); | |
111 | end_pfn = min_t(unsigned long, end_pfn, | |
112 | domain->geometry.aperture_end >> order); | |
113 | } | |
114 | ||
115 | /* All we can safely do with an existing domain is enlarge it */ | |
116 | if (iovad->start_pfn) { | |
117 | if (1UL << order != iovad->granule || | |
118 | base_pfn != iovad->start_pfn || | |
119 | end_pfn < iovad->dma_32bit_pfn) { | |
120 | pr_warn("Incompatible range for DMA domain\n"); | |
121 | return -EFAULT; | |
122 | } | |
123 | iovad->dma_32bit_pfn = end_pfn; | |
124 | } else { | |
125 | init_iova_domain(iovad, 1UL << order, base_pfn, end_pfn); | |
126 | } | |
127 | return 0; | |
128 | } | |
129 | EXPORT_SYMBOL(iommu_dma_init_domain); | |
130 | ||
131 | /** | |
132 | * dma_direction_to_prot - Translate DMA API directions to IOMMU API page flags | |
133 | * @dir: Direction of DMA transfer | |
134 | * @coherent: Is the DMA master cache-coherent? | |
135 | * | |
136 | * Return: corresponding IOMMU API page protection flags | |
137 | */ | |
138 | int dma_direction_to_prot(enum dma_data_direction dir, bool coherent) | |
139 | { | |
140 | int prot = coherent ? IOMMU_CACHE : 0; | |
141 | ||
142 | switch (dir) { | |
143 | case DMA_BIDIRECTIONAL: | |
144 | return prot | IOMMU_READ | IOMMU_WRITE; | |
145 | case DMA_TO_DEVICE: | |
146 | return prot | IOMMU_READ; | |
147 | case DMA_FROM_DEVICE: | |
148 | return prot | IOMMU_WRITE; | |
149 | default: | |
150 | return 0; | |
151 | } | |
152 | } | |
153 | ||
154 | static struct iova *__alloc_iova(struct iova_domain *iovad, size_t size, | |
155 | dma_addr_t dma_limit) | |
156 | { | |
157 | unsigned long shift = iova_shift(iovad); | |
158 | unsigned long length = iova_align(iovad, size) >> shift; | |
159 | ||
160 | /* | |
161 | * Enforce size-alignment to be safe - there could perhaps be an | |
162 | * attribute to control this per-device, or at least per-domain... | |
163 | */ | |
164 | return alloc_iova(iovad, length, dma_limit >> shift, true); | |
165 | } | |
166 | ||
167 | /* The IOVA allocator knows what we mapped, so just unmap whatever that was */ | |
168 | static void __iommu_dma_unmap(struct iommu_domain *domain, dma_addr_t dma_addr) | |
169 | { | |
170 | struct iova_domain *iovad = domain->iova_cookie; | |
171 | unsigned long shift = iova_shift(iovad); | |
172 | unsigned long pfn = dma_addr >> shift; | |
173 | struct iova *iova = find_iova(iovad, pfn); | |
174 | size_t size; | |
175 | ||
176 | if (WARN_ON(!iova)) | |
177 | return; | |
178 | ||
179 | size = iova_size(iova) << shift; | |
180 | size -= iommu_unmap(domain, pfn << shift, size); | |
181 | /* ...and if we can't, then something is horribly, horribly wrong */ | |
182 | WARN_ON(size > 0); | |
183 | __free_iova(iovad, iova); | |
184 | } | |
185 | ||
186 | static void __iommu_dma_free_pages(struct page **pages, int count) | |
187 | { | |
188 | while (count--) | |
189 | __free_page(pages[count]); | |
190 | kvfree(pages); | |
191 | } | |
192 | ||
193 | static struct page **__iommu_dma_alloc_pages(unsigned int count, gfp_t gfp) | |
194 | { | |
195 | struct page **pages; | |
196 | unsigned int i = 0, array_size = count * sizeof(*pages); | |
0a9afeda | 197 | unsigned int order = MAX_ORDER; |
0db2e5d1 RM |
198 | |
199 | if (array_size <= PAGE_SIZE) | |
200 | pages = kzalloc(array_size, GFP_KERNEL); | |
201 | else | |
202 | pages = vzalloc(array_size); | |
203 | if (!pages) | |
204 | return NULL; | |
205 | ||
206 | /* IOMMU can map any pages, so himem can also be used here */ | |
207 | gfp |= __GFP_NOWARN | __GFP_HIGHMEM; | |
208 | ||
209 | while (count) { | |
210 | struct page *page = NULL; | |
0a9afeda | 211 | int j; |
0db2e5d1 RM |
212 | |
213 | /* | |
214 | * Higher-order allocations are a convenience rather | |
215 | * than a necessity, hence using __GFP_NORETRY until | |
216 | * falling back to single-page allocations. | |
217 | */ | |
0a9afeda RM |
218 | for (order = min_t(unsigned int, order, __fls(count)); |
219 | order > 0; order--) { | |
0db2e5d1 RM |
220 | page = alloc_pages(gfp | __GFP_NORETRY, order); |
221 | if (!page) | |
222 | continue; | |
223 | if (PageCompound(page)) { | |
224 | if (!split_huge_page(page)) | |
225 | break; | |
226 | __free_pages(page, order); | |
227 | } else { | |
228 | split_page(page, order); | |
229 | break; | |
230 | } | |
231 | } | |
232 | if (!page) | |
233 | page = alloc_page(gfp); | |
234 | if (!page) { | |
235 | __iommu_dma_free_pages(pages, i); | |
236 | return NULL; | |
237 | } | |
238 | j = 1 << order; | |
239 | count -= j; | |
240 | while (j--) | |
241 | pages[i++] = page++; | |
242 | } | |
243 | return pages; | |
244 | } | |
245 | ||
246 | /** | |
247 | * iommu_dma_free - Free a buffer allocated by iommu_dma_alloc() | |
248 | * @dev: Device which owns this buffer | |
249 | * @pages: Array of buffer pages as returned by iommu_dma_alloc() | |
250 | * @size: Size of buffer in bytes | |
251 | * @handle: DMA address of buffer | |
252 | * | |
253 | * Frees both the pages associated with the buffer, and the array | |
254 | * describing them | |
255 | */ | |
256 | void iommu_dma_free(struct device *dev, struct page **pages, size_t size, | |
257 | dma_addr_t *handle) | |
258 | { | |
259 | __iommu_dma_unmap(iommu_get_domain_for_dev(dev), *handle); | |
260 | __iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT); | |
261 | *handle = DMA_ERROR_CODE; | |
262 | } | |
263 | ||
264 | /** | |
265 | * iommu_dma_alloc - Allocate and map a buffer contiguous in IOVA space | |
266 | * @dev: Device to allocate memory for. Must be a real device | |
267 | * attached to an iommu_dma_domain | |
268 | * @size: Size of buffer in bytes | |
269 | * @gfp: Allocation flags | |
270 | * @prot: IOMMU mapping flags | |
271 | * @handle: Out argument for allocated DMA handle | |
272 | * @flush_page: Arch callback which must ensure PAGE_SIZE bytes from the | |
273 | * given VA/PA are visible to the given non-coherent device. | |
274 | * | |
275 | * If @size is less than PAGE_SIZE, then a full CPU page will be allocated, | |
276 | * but an IOMMU which supports smaller pages might not map the whole thing. | |
277 | * | |
278 | * Return: Array of struct page pointers describing the buffer, | |
279 | * or NULL on failure. | |
280 | */ | |
281 | struct page **iommu_dma_alloc(struct device *dev, size_t size, | |
282 | gfp_t gfp, int prot, dma_addr_t *handle, | |
283 | void (*flush_page)(struct device *, const void *, phys_addr_t)) | |
284 | { | |
285 | struct iommu_domain *domain = iommu_get_domain_for_dev(dev); | |
286 | struct iova_domain *iovad = domain->iova_cookie; | |
287 | struct iova *iova; | |
288 | struct page **pages; | |
289 | struct sg_table sgt; | |
290 | dma_addr_t dma_addr; | |
291 | unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT; | |
292 | ||
293 | *handle = DMA_ERROR_CODE; | |
294 | ||
295 | pages = __iommu_dma_alloc_pages(count, gfp); | |
296 | if (!pages) | |
297 | return NULL; | |
298 | ||
299 | iova = __alloc_iova(iovad, size, dev->coherent_dma_mask); | |
300 | if (!iova) | |
301 | goto out_free_pages; | |
302 | ||
303 | size = iova_align(iovad, size); | |
304 | if (sg_alloc_table_from_pages(&sgt, pages, count, 0, size, GFP_KERNEL)) | |
305 | goto out_free_iova; | |
306 | ||
307 | if (!(prot & IOMMU_CACHE)) { | |
308 | struct sg_mapping_iter miter; | |
309 | /* | |
310 | * The CPU-centric flushing implied by SG_MITER_TO_SG isn't | |
311 | * sufficient here, so skip it by using the "wrong" direction. | |
312 | */ | |
313 | sg_miter_start(&miter, sgt.sgl, sgt.orig_nents, SG_MITER_FROM_SG); | |
314 | while (sg_miter_next(&miter)) | |
315 | flush_page(dev, miter.addr, page_to_phys(miter.page)); | |
316 | sg_miter_stop(&miter); | |
317 | } | |
318 | ||
319 | dma_addr = iova_dma_addr(iovad, iova); | |
320 | if (iommu_map_sg(domain, dma_addr, sgt.sgl, sgt.orig_nents, prot) | |
321 | < size) | |
322 | goto out_free_sg; | |
323 | ||
324 | *handle = dma_addr; | |
325 | sg_free_table(&sgt); | |
326 | return pages; | |
327 | ||
328 | out_free_sg: | |
329 | sg_free_table(&sgt); | |
330 | out_free_iova: | |
331 | __free_iova(iovad, iova); | |
332 | out_free_pages: | |
333 | __iommu_dma_free_pages(pages, count); | |
334 | return NULL; | |
335 | } | |
336 | ||
337 | /** | |
338 | * iommu_dma_mmap - Map a buffer into provided user VMA | |
339 | * @pages: Array representing buffer from iommu_dma_alloc() | |
340 | * @size: Size of buffer in bytes | |
341 | * @vma: VMA describing requested userspace mapping | |
342 | * | |
343 | * Maps the pages of the buffer in @pages into @vma. The caller is responsible | |
344 | * for verifying the correct size and protection of @vma beforehand. | |
345 | */ | |
346 | ||
347 | int iommu_dma_mmap(struct page **pages, size_t size, struct vm_area_struct *vma) | |
348 | { | |
349 | unsigned long uaddr = vma->vm_start; | |
350 | unsigned int i, count = PAGE_ALIGN(size) >> PAGE_SHIFT; | |
351 | int ret = -ENXIO; | |
352 | ||
353 | for (i = vma->vm_pgoff; i < count && uaddr < vma->vm_end; i++) { | |
354 | ret = vm_insert_page(vma, uaddr, pages[i]); | |
355 | if (ret) | |
356 | break; | |
357 | uaddr += PAGE_SIZE; | |
358 | } | |
359 | return ret; | |
360 | } | |
361 | ||
362 | dma_addr_t iommu_dma_map_page(struct device *dev, struct page *page, | |
363 | unsigned long offset, size_t size, int prot) | |
364 | { | |
365 | dma_addr_t dma_addr; | |
366 | struct iommu_domain *domain = iommu_get_domain_for_dev(dev); | |
367 | struct iova_domain *iovad = domain->iova_cookie; | |
368 | phys_addr_t phys = page_to_phys(page) + offset; | |
369 | size_t iova_off = iova_offset(iovad, phys); | |
370 | size_t len = iova_align(iovad, size + iova_off); | |
371 | struct iova *iova = __alloc_iova(iovad, len, dma_get_mask(dev)); | |
372 | ||
373 | if (!iova) | |
374 | return DMA_ERROR_CODE; | |
375 | ||
376 | dma_addr = iova_dma_addr(iovad, iova); | |
377 | if (iommu_map(domain, dma_addr, phys - iova_off, len, prot)) { | |
378 | __free_iova(iovad, iova); | |
379 | return DMA_ERROR_CODE; | |
380 | } | |
381 | return dma_addr + iova_off; | |
382 | } | |
383 | ||
384 | void iommu_dma_unmap_page(struct device *dev, dma_addr_t handle, size_t size, | |
385 | enum dma_data_direction dir, struct dma_attrs *attrs) | |
386 | { | |
387 | __iommu_dma_unmap(iommu_get_domain_for_dev(dev), handle); | |
388 | } | |
389 | ||
390 | /* | |
391 | * Prepare a successfully-mapped scatterlist to give back to the caller. | |
392 | * Handling IOVA concatenation can come later, if needed | |
393 | */ | |
394 | static int __finalise_sg(struct device *dev, struct scatterlist *sg, int nents, | |
395 | dma_addr_t dma_addr) | |
396 | { | |
397 | struct scatterlist *s; | |
398 | int i; | |
399 | ||
400 | for_each_sg(sg, s, nents, i) { | |
401 | /* Un-swizzling the fields here, hence the naming mismatch */ | |
402 | unsigned int s_offset = sg_dma_address(s); | |
403 | unsigned int s_length = sg_dma_len(s); | |
404 | unsigned int s_dma_len = s->length; | |
405 | ||
406 | s->offset = s_offset; | |
407 | s->length = s_length; | |
408 | sg_dma_address(s) = dma_addr + s_offset; | |
409 | dma_addr += s_dma_len; | |
410 | } | |
411 | return i; | |
412 | } | |
413 | ||
414 | /* | |
415 | * If mapping failed, then just restore the original list, | |
416 | * but making sure the DMA fields are invalidated. | |
417 | */ | |
418 | static void __invalidate_sg(struct scatterlist *sg, int nents) | |
419 | { | |
420 | struct scatterlist *s; | |
421 | int i; | |
422 | ||
423 | for_each_sg(sg, s, nents, i) { | |
424 | if (sg_dma_address(s) != DMA_ERROR_CODE) | |
425 | s->offset = sg_dma_address(s); | |
426 | if (sg_dma_len(s)) | |
427 | s->length = sg_dma_len(s); | |
428 | sg_dma_address(s) = DMA_ERROR_CODE; | |
429 | sg_dma_len(s) = 0; | |
430 | } | |
431 | } | |
432 | ||
433 | /* | |
434 | * The DMA API client is passing in a scatterlist which could describe | |
435 | * any old buffer layout, but the IOMMU API requires everything to be | |
436 | * aligned to IOMMU pages. Hence the need for this complicated bit of | |
437 | * impedance-matching, to be able to hand off a suitably-aligned list, | |
438 | * but still preserve the original offsets and sizes for the caller. | |
439 | */ | |
440 | int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg, | |
441 | int nents, int prot) | |
442 | { | |
443 | struct iommu_domain *domain = iommu_get_domain_for_dev(dev); | |
444 | struct iova_domain *iovad = domain->iova_cookie; | |
445 | struct iova *iova; | |
446 | struct scatterlist *s, *prev = NULL; | |
447 | dma_addr_t dma_addr; | |
448 | size_t iova_len = 0; | |
449 | int i; | |
450 | ||
451 | /* | |
452 | * Work out how much IOVA space we need, and align the segments to | |
453 | * IOVA granules for the IOMMU driver to handle. With some clever | |
454 | * trickery we can modify the list in-place, but reversibly, by | |
455 | * hiding the original data in the as-yet-unused DMA fields. | |
456 | */ | |
457 | for_each_sg(sg, s, nents, i) { | |
458 | size_t s_offset = iova_offset(iovad, s->offset); | |
459 | size_t s_length = s->length; | |
460 | ||
164afb1d | 461 | sg_dma_address(s) = s_offset; |
0db2e5d1 RM |
462 | sg_dma_len(s) = s_length; |
463 | s->offset -= s_offset; | |
464 | s_length = iova_align(iovad, s_length + s_offset); | |
465 | s->length = s_length; | |
466 | ||
467 | /* | |
468 | * The simple way to avoid the rare case of a segment | |
469 | * crossing the boundary mask is to pad the previous one | |
470 | * to end at a naturally-aligned IOVA for this one's size, | |
471 | * at the cost of potentially over-allocating a little. | |
472 | */ | |
473 | if (prev) { | |
474 | size_t pad_len = roundup_pow_of_two(s_length); | |
475 | ||
476 | pad_len = (pad_len - iova_len) & (pad_len - 1); | |
477 | prev->length += pad_len; | |
478 | iova_len += pad_len; | |
479 | } | |
480 | ||
481 | iova_len += s_length; | |
482 | prev = s; | |
483 | } | |
484 | ||
485 | iova = __alloc_iova(iovad, iova_len, dma_get_mask(dev)); | |
486 | if (!iova) | |
487 | goto out_restore_sg; | |
488 | ||
489 | /* | |
490 | * We'll leave any physical concatenation to the IOMMU driver's | |
491 | * implementation - it knows better than we do. | |
492 | */ | |
493 | dma_addr = iova_dma_addr(iovad, iova); | |
494 | if (iommu_map_sg(domain, dma_addr, sg, nents, prot) < iova_len) | |
495 | goto out_free_iova; | |
496 | ||
497 | return __finalise_sg(dev, sg, nents, dma_addr); | |
498 | ||
499 | out_free_iova: | |
500 | __free_iova(iovad, iova); | |
501 | out_restore_sg: | |
502 | __invalidate_sg(sg, nents); | |
503 | return 0; | |
504 | } | |
505 | ||
506 | void iommu_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, | |
507 | enum dma_data_direction dir, struct dma_attrs *attrs) | |
508 | { | |
509 | /* | |
510 | * The scatterlist segments are mapped into a single | |
511 | * contiguous IOVA allocation, so this is incredibly easy. | |
512 | */ | |
513 | __iommu_dma_unmap(iommu_get_domain_for_dev(dev), sg_dma_address(sg)); | |
514 | } | |
515 | ||
516 | int iommu_dma_supported(struct device *dev, u64 mask) | |
517 | { | |
518 | /* | |
519 | * 'Special' IOMMUs which don't have the same addressing capability | |
520 | * as the CPU will have to wait until we have some way to query that | |
521 | * before they'll be able to use this framework. | |
522 | */ | |
523 | return 1; | |
524 | } | |
525 | ||
526 | int iommu_dma_mapping_error(struct device *dev, dma_addr_t dma_addr) | |
527 | { | |
528 | return dma_addr == DMA_ERROR_CODE; | |
529 | } |