Loading...
1/*
2 * linux/arch/arm/mm/dma-mapping.c
3 *
4 * Copyright (C) 2000-2004 Russell King
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
9 *
10 * DMA uncached mapping support.
11 */
12#include <linux/module.h>
13#include <linux/mm.h>
14#include <linux/gfp.h>
15#include <linux/errno.h>
16#include <linux/list.h>
17#include <linux/init.h>
18#include <linux/device.h>
19#include <linux/dma-mapping.h>
20#include <linux/highmem.h>
21
22#include <asm/memory.h>
23#include <asm/highmem.h>
24#include <asm/cacheflush.h>
25#include <asm/tlbflush.h>
26#include <asm/sizes.h>
27
28#include "mm.h"
29
30static u64 get_coherent_dma_mask(struct device *dev)
31{
32 u64 mask = (u64)arm_dma_limit;
33
34 if (dev) {
35 mask = dev->coherent_dma_mask;
36
37 /*
38 * Sanity check the DMA mask - it must be non-zero, and
39 * must be able to be satisfied by a DMA allocation.
40 */
41 if (mask == 0) {
42 dev_warn(dev, "coherent DMA mask is unset\n");
43 return 0;
44 }
45
46 if ((~mask) & (u64)arm_dma_limit) {
47 dev_warn(dev, "coherent DMA mask %#llx is smaller "
48 "than system GFP_DMA mask %#llx\n",
49 mask, (u64)arm_dma_limit);
50 return 0;
51 }
52 }
53
54 return mask;
55}
56
57/*
58 * Allocate a DMA buffer for 'dev' of size 'size' using the
59 * specified gfp mask. Note that 'size' must be page aligned.
60 */
61static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp)
62{
63 unsigned long order = get_order(size);
64 struct page *page, *p, *e;
65 void *ptr;
66 u64 mask = get_coherent_dma_mask(dev);
67
68#ifdef CONFIG_DMA_API_DEBUG
69 u64 limit = (mask + 1) & ~mask;
70 if (limit && size >= limit) {
71 dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n",
72 size, mask);
73 return NULL;
74 }
75#endif
76
77 if (!mask)
78 return NULL;
79
80 if (mask < 0xffffffffULL)
81 gfp |= GFP_DMA;
82
83 page = alloc_pages(gfp, order);
84 if (!page)
85 return NULL;
86
87 /*
88 * Now split the huge page and free the excess pages
89 */
90 split_page(page, order);
91 for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
92 __free_page(p);
93
94 /*
95 * Ensure that the allocated pages are zeroed, and that any data
96 * lurking in the kernel direct-mapped region is invalidated.
97 */
98 ptr = page_address(page);
99 memset(ptr, 0, size);
100 dmac_flush_range(ptr, ptr + size);
101 outer_flush_range(__pa(ptr), __pa(ptr) + size);
102
103 return page;
104}
105
106/*
107 * Free a DMA buffer. 'size' must be page aligned.
108 */
109static void __dma_free_buffer(struct page *page, size_t size)
110{
111 struct page *e = page + (size >> PAGE_SHIFT);
112
113 while (page < e) {
114 __free_page(page);
115 page++;
116 }
117}
118
119#ifdef CONFIG_MMU
120/* Sanity check size */
121#if (CONSISTENT_DMA_SIZE % SZ_2M)
122#error "CONSISTENT_DMA_SIZE must be multiple of 2MiB"
123#endif
124
125#define CONSISTENT_OFFSET(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
126#define CONSISTENT_PTE_INDEX(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PGDIR_SHIFT)
127#define NUM_CONSISTENT_PTES (CONSISTENT_DMA_SIZE >> PGDIR_SHIFT)
128
129/*
130 * These are the page tables (2MB each) covering uncached, DMA consistent allocations
131 */
132static pte_t *consistent_pte[NUM_CONSISTENT_PTES];
133
134#include "vmregion.h"
135
136static struct arm_vmregion_head consistent_head = {
137 .vm_lock = __SPIN_LOCK_UNLOCKED(&consistent_head.vm_lock),
138 .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
139 .vm_start = CONSISTENT_BASE,
140 .vm_end = CONSISTENT_END,
141};
142
143#ifdef CONFIG_HUGETLB_PAGE
144#error ARM Coherent DMA allocator does not (yet) support huge TLB
145#endif
146
147/*
148 * Initialise the consistent memory allocation.
149 */
150static int __init consistent_init(void)
151{
152 int ret = 0;
153 pgd_t *pgd;
154 pud_t *pud;
155 pmd_t *pmd;
156 pte_t *pte;
157 int i = 0;
158 u32 base = CONSISTENT_BASE;
159
160 do {
161 pgd = pgd_offset(&init_mm, base);
162
163 pud = pud_alloc(&init_mm, pgd, base);
164 if (!pud) {
165 printk(KERN_ERR "%s: no pud tables\n", __func__);
166 ret = -ENOMEM;
167 break;
168 }
169
170 pmd = pmd_alloc(&init_mm, pud, base);
171 if (!pmd) {
172 printk(KERN_ERR "%s: no pmd tables\n", __func__);
173 ret = -ENOMEM;
174 break;
175 }
176 WARN_ON(!pmd_none(*pmd));
177
178 pte = pte_alloc_kernel(pmd, base);
179 if (!pte) {
180 printk(KERN_ERR "%s: no pte tables\n", __func__);
181 ret = -ENOMEM;
182 break;
183 }
184
185 consistent_pte[i++] = pte;
186 base += (1 << PGDIR_SHIFT);
187 } while (base < CONSISTENT_END);
188
189 return ret;
190}
191
192core_initcall(consistent_init);
193
194static void *
195__dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot)
196{
197 struct arm_vmregion *c;
198 size_t align;
199 int bit;
200
201 if (!consistent_pte[0]) {
202 printk(KERN_ERR "%s: not initialised\n", __func__);
203 dump_stack();
204 return NULL;
205 }
206
207 /*
208 * Align the virtual region allocation - maximum alignment is
209 * a section size, minimum is a page size. This helps reduce
210 * fragmentation of the DMA space, and also prevents allocations
211 * smaller than a section from crossing a section boundary.
212 */
213 bit = fls(size - 1);
214 if (bit > SECTION_SHIFT)
215 bit = SECTION_SHIFT;
216 align = 1 << bit;
217
218 /*
219 * Allocate a virtual address in the consistent mapping region.
220 */
221 c = arm_vmregion_alloc(&consistent_head, align, size,
222 gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
223 if (c) {
224 pte_t *pte;
225 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
226 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
227
228 pte = consistent_pte[idx] + off;
229 c->vm_pages = page;
230
231 do {
232 BUG_ON(!pte_none(*pte));
233
234 set_pte_ext(pte, mk_pte(page, prot), 0);
235 page++;
236 pte++;
237 off++;
238 if (off >= PTRS_PER_PTE) {
239 off = 0;
240 pte = consistent_pte[++idx];
241 }
242 } while (size -= PAGE_SIZE);
243
244 dsb();
245
246 return (void *)c->vm_start;
247 }
248 return NULL;
249}
250
251static void __dma_free_remap(void *cpu_addr, size_t size)
252{
253 struct arm_vmregion *c;
254 unsigned long addr;
255 pte_t *ptep;
256 int idx;
257 u32 off;
258
259 c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
260 if (!c) {
261 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
262 __func__, cpu_addr);
263 dump_stack();
264 return;
265 }
266
267 if ((c->vm_end - c->vm_start) != size) {
268 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
269 __func__, c->vm_end - c->vm_start, size);
270 dump_stack();
271 size = c->vm_end - c->vm_start;
272 }
273
274 idx = CONSISTENT_PTE_INDEX(c->vm_start);
275 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
276 ptep = consistent_pte[idx] + off;
277 addr = c->vm_start;
278 do {
279 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
280
281 ptep++;
282 addr += PAGE_SIZE;
283 off++;
284 if (off >= PTRS_PER_PTE) {
285 off = 0;
286 ptep = consistent_pte[++idx];
287 }
288
289 if (pte_none(pte) || !pte_present(pte))
290 printk(KERN_CRIT "%s: bad page in kernel page table\n",
291 __func__);
292 } while (size -= PAGE_SIZE);
293
294 flush_tlb_kernel_range(c->vm_start, c->vm_end);
295
296 arm_vmregion_free(&consistent_head, c);
297}
298
299#else /* !CONFIG_MMU */
300
301#define __dma_alloc_remap(page, size, gfp, prot) page_address(page)
302#define __dma_free_remap(addr, size) do { } while (0)
303
304#endif /* CONFIG_MMU */
305
306static void *
307__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
308 pgprot_t prot)
309{
310 struct page *page;
311 void *addr;
312
313 *handle = ~0;
314 size = PAGE_ALIGN(size);
315
316 page = __dma_alloc_buffer(dev, size, gfp);
317 if (!page)
318 return NULL;
319
320 if (!arch_is_coherent())
321 addr = __dma_alloc_remap(page, size, gfp, prot);
322 else
323 addr = page_address(page);
324
325 if (addr)
326 *handle = pfn_to_dma(dev, page_to_pfn(page));
327 else
328 __dma_free_buffer(page, size);
329
330 return addr;
331}
332
333/*
334 * Allocate DMA-coherent memory space and return both the kernel remapped
335 * virtual and bus address for that space.
336 */
337void *
338dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
339{
340 void *memory;
341
342 if (dma_alloc_from_coherent(dev, size, handle, &memory))
343 return memory;
344
345 return __dma_alloc(dev, size, handle, gfp,
346 pgprot_dmacoherent(pgprot_kernel));
347}
348EXPORT_SYMBOL(dma_alloc_coherent);
349
350/*
351 * Allocate a writecombining region, in much the same way as
352 * dma_alloc_coherent above.
353 */
354void *
355dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
356{
357 return __dma_alloc(dev, size, handle, gfp,
358 pgprot_writecombine(pgprot_kernel));
359}
360EXPORT_SYMBOL(dma_alloc_writecombine);
361
362static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
363 void *cpu_addr, dma_addr_t dma_addr, size_t size)
364{
365 int ret = -ENXIO;
366#ifdef CONFIG_MMU
367 unsigned long user_size, kern_size;
368 struct arm_vmregion *c;
369
370 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
371
372 c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
373 if (c) {
374 unsigned long off = vma->vm_pgoff;
375
376 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
377
378 if (off < kern_size &&
379 user_size <= (kern_size - off)) {
380 ret = remap_pfn_range(vma, vma->vm_start,
381 page_to_pfn(c->vm_pages) + off,
382 user_size << PAGE_SHIFT,
383 vma->vm_page_prot);
384 }
385 }
386#endif /* CONFIG_MMU */
387
388 return ret;
389}
390
391int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
392 void *cpu_addr, dma_addr_t dma_addr, size_t size)
393{
394 vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot);
395 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
396}
397EXPORT_SYMBOL(dma_mmap_coherent);
398
399int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
400 void *cpu_addr, dma_addr_t dma_addr, size_t size)
401{
402 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
403 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
404}
405EXPORT_SYMBOL(dma_mmap_writecombine);
406
407/*
408 * free a page as defined by the above mapping.
409 * Must not be called with IRQs disabled.
410 */
411void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
412{
413 WARN_ON(irqs_disabled());
414
415 if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
416 return;
417
418 size = PAGE_ALIGN(size);
419
420 if (!arch_is_coherent())
421 __dma_free_remap(cpu_addr, size);
422
423 __dma_free_buffer(pfn_to_page(dma_to_pfn(dev, handle)), size);
424}
425EXPORT_SYMBOL(dma_free_coherent);
426
427/*
428 * Make an area consistent for devices.
429 * Note: Drivers should NOT use this function directly, as it will break
430 * platforms with CONFIG_DMABOUNCE.
431 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
432 */
433void ___dma_single_cpu_to_dev(const void *kaddr, size_t size,
434 enum dma_data_direction dir)
435{
436 unsigned long paddr;
437
438 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
439
440 dmac_map_area(kaddr, size, dir);
441
442 paddr = __pa(kaddr);
443 if (dir == DMA_FROM_DEVICE) {
444 outer_inv_range(paddr, paddr + size);
445 } else {
446 outer_clean_range(paddr, paddr + size);
447 }
448 /* FIXME: non-speculating: flush on bidirectional mappings? */
449}
450EXPORT_SYMBOL(___dma_single_cpu_to_dev);
451
452void ___dma_single_dev_to_cpu(const void *kaddr, size_t size,
453 enum dma_data_direction dir)
454{
455 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
456
457 /* FIXME: non-speculating: not required */
458 /* don't bother invalidating if DMA to device */
459 if (dir != DMA_TO_DEVICE) {
460 unsigned long paddr = __pa(kaddr);
461 outer_inv_range(paddr, paddr + size);
462 }
463
464 dmac_unmap_area(kaddr, size, dir);
465}
466EXPORT_SYMBOL(___dma_single_dev_to_cpu);
467
468static void dma_cache_maint_page(struct page *page, unsigned long offset,
469 size_t size, enum dma_data_direction dir,
470 void (*op)(const void *, size_t, int))
471{
472 /*
473 * A single sg entry may refer to multiple physically contiguous
474 * pages. But we still need to process highmem pages individually.
475 * If highmem is not configured then the bulk of this loop gets
476 * optimized out.
477 */
478 size_t left = size;
479 do {
480 size_t len = left;
481 void *vaddr;
482
483 if (PageHighMem(page)) {
484 if (len + offset > PAGE_SIZE) {
485 if (offset >= PAGE_SIZE) {
486 page += offset / PAGE_SIZE;
487 offset %= PAGE_SIZE;
488 }
489 len = PAGE_SIZE - offset;
490 }
491 vaddr = kmap_high_get(page);
492 if (vaddr) {
493 vaddr += offset;
494 op(vaddr, len, dir);
495 kunmap_high(page);
496 } else if (cache_is_vipt()) {
497 /* unmapped pages might still be cached */
498 vaddr = kmap_atomic(page);
499 op(vaddr + offset, len, dir);
500 kunmap_atomic(vaddr);
501 }
502 } else {
503 vaddr = page_address(page) + offset;
504 op(vaddr, len, dir);
505 }
506 offset = 0;
507 page++;
508 left -= len;
509 } while (left);
510}
511
512void ___dma_page_cpu_to_dev(struct page *page, unsigned long off,
513 size_t size, enum dma_data_direction dir)
514{
515 unsigned long paddr;
516
517 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
518
519 paddr = page_to_phys(page) + off;
520 if (dir == DMA_FROM_DEVICE) {
521 outer_inv_range(paddr, paddr + size);
522 } else {
523 outer_clean_range(paddr, paddr + size);
524 }
525 /* FIXME: non-speculating: flush on bidirectional mappings? */
526}
527EXPORT_SYMBOL(___dma_page_cpu_to_dev);
528
529void ___dma_page_dev_to_cpu(struct page *page, unsigned long off,
530 size_t size, enum dma_data_direction dir)
531{
532 unsigned long paddr = page_to_phys(page) + off;
533
534 /* FIXME: non-speculating: not required */
535 /* don't bother invalidating if DMA to device */
536 if (dir != DMA_TO_DEVICE)
537 outer_inv_range(paddr, paddr + size);
538
539 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
540
541 /*
542 * Mark the D-cache clean for this page to avoid extra flushing.
543 */
544 if (dir != DMA_TO_DEVICE && off == 0 && size >= PAGE_SIZE)
545 set_bit(PG_dcache_clean, &page->flags);
546}
547EXPORT_SYMBOL(___dma_page_dev_to_cpu);
548
549/**
550 * dma_map_sg - map a set of SG buffers for streaming mode DMA
551 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
552 * @sg: list of buffers
553 * @nents: number of buffers to map
554 * @dir: DMA transfer direction
555 *
556 * Map a set of buffers described by scatterlist in streaming mode for DMA.
557 * This is the scatter-gather version of the dma_map_single interface.
558 * Here the scatter gather list elements are each tagged with the
559 * appropriate dma address and length. They are obtained via
560 * sg_dma_{address,length}.
561 *
562 * Device ownership issues as mentioned for dma_map_single are the same
563 * here.
564 */
565int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
566 enum dma_data_direction dir)
567{
568 struct scatterlist *s;
569 int i, j;
570
571 BUG_ON(!valid_dma_direction(dir));
572
573 for_each_sg(sg, s, nents, i) {
574 s->dma_address = __dma_map_page(dev, sg_page(s), s->offset,
575 s->length, dir);
576 if (dma_mapping_error(dev, s->dma_address))
577 goto bad_mapping;
578 }
579 debug_dma_map_sg(dev, sg, nents, nents, dir);
580 return nents;
581
582 bad_mapping:
583 for_each_sg(sg, s, i, j)
584 __dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
585 return 0;
586}
587EXPORT_SYMBOL(dma_map_sg);
588
589/**
590 * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
591 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
592 * @sg: list of buffers
593 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
594 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
595 *
596 * Unmap a set of streaming mode DMA translations. Again, CPU access
597 * rules concerning calls here are the same as for dma_unmap_single().
598 */
599void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
600 enum dma_data_direction dir)
601{
602 struct scatterlist *s;
603 int i;
604
605 debug_dma_unmap_sg(dev, sg, nents, dir);
606
607 for_each_sg(sg, s, nents, i)
608 __dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
609}
610EXPORT_SYMBOL(dma_unmap_sg);
611
612/**
613 * dma_sync_sg_for_cpu
614 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
615 * @sg: list of buffers
616 * @nents: number of buffers to map (returned from dma_map_sg)
617 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
618 */
619void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
620 int nents, enum dma_data_direction dir)
621{
622 struct scatterlist *s;
623 int i;
624
625 for_each_sg(sg, s, nents, i) {
626 if (!dmabounce_sync_for_cpu(dev, sg_dma_address(s), 0,
627 sg_dma_len(s), dir))
628 continue;
629
630 __dma_page_dev_to_cpu(sg_page(s), s->offset,
631 s->length, dir);
632 }
633
634 debug_dma_sync_sg_for_cpu(dev, sg, nents, dir);
635}
636EXPORT_SYMBOL(dma_sync_sg_for_cpu);
637
638/**
639 * dma_sync_sg_for_device
640 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
641 * @sg: list of buffers
642 * @nents: number of buffers to map (returned from dma_map_sg)
643 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
644 */
645void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
646 int nents, enum dma_data_direction dir)
647{
648 struct scatterlist *s;
649 int i;
650
651 for_each_sg(sg, s, nents, i) {
652 if (!dmabounce_sync_for_device(dev, sg_dma_address(s), 0,
653 sg_dma_len(s), dir))
654 continue;
655
656 __dma_page_cpu_to_dev(sg_page(s), s->offset,
657 s->length, dir);
658 }
659
660 debug_dma_sync_sg_for_device(dev, sg, nents, dir);
661}
662EXPORT_SYMBOL(dma_sync_sg_for_device);
663
664/*
665 * Return whether the given device DMA address mask can be supported
666 * properly. For example, if your device can only drive the low 24-bits
667 * during bus mastering, then you would pass 0x00ffffff as the mask
668 * to this function.
669 */
670int dma_supported(struct device *dev, u64 mask)
671{
672 if (mask < (u64)arm_dma_limit)
673 return 0;
674 return 1;
675}
676EXPORT_SYMBOL(dma_supported);
677
678int dma_set_mask(struct device *dev, u64 dma_mask)
679{
680 if (!dev->dma_mask || !dma_supported(dev, dma_mask))
681 return -EIO;
682
683#ifndef CONFIG_DMABOUNCE
684 *dev->dma_mask = dma_mask;
685#endif
686
687 return 0;
688}
689EXPORT_SYMBOL(dma_set_mask);
690
691#define PREALLOC_DMA_DEBUG_ENTRIES 4096
692
693static int __init dma_debug_do_init(void)
694{
695 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
696 return 0;
697}
698fs_initcall(dma_debug_do_init);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/arch/arm/mm/dma-mapping.c
4 *
5 * Copyright (C) 2000-2004 Russell King
6 *
7 * DMA uncached mapping support.
8 */
9#include <linux/module.h>
10#include <linux/mm.h>
11#include <linux/genalloc.h>
12#include <linux/gfp.h>
13#include <linux/errno.h>
14#include <linux/list.h>
15#include <linux/init.h>
16#include <linux/device.h>
17#include <linux/dma-direct.h>
18#include <linux/dma-map-ops.h>
19#include <linux/highmem.h>
20#include <linux/memblock.h>
21#include <linux/slab.h>
22#include <linux/iommu.h>
23#include <linux/io.h>
24#include <linux/vmalloc.h>
25#include <linux/sizes.h>
26#include <linux/cma.h>
27
28#include <asm/memory.h>
29#include <asm/highmem.h>
30#include <asm/cacheflush.h>
31#include <asm/tlbflush.h>
32#include <asm/mach/arch.h>
33#include <asm/dma-iommu.h>
34#include <asm/mach/map.h>
35#include <asm/system_info.h>
36#include <xen/swiotlb-xen.h>
37
38#include "dma.h"
39#include "mm.h"
40
41struct arm_dma_alloc_args {
42 struct device *dev;
43 size_t size;
44 gfp_t gfp;
45 pgprot_t prot;
46 const void *caller;
47 bool want_vaddr;
48 int coherent_flag;
49};
50
51struct arm_dma_free_args {
52 struct device *dev;
53 size_t size;
54 void *cpu_addr;
55 struct page *page;
56 bool want_vaddr;
57};
58
59#define NORMAL 0
60#define COHERENT 1
61
62struct arm_dma_allocator {
63 void *(*alloc)(struct arm_dma_alloc_args *args,
64 struct page **ret_page);
65 void (*free)(struct arm_dma_free_args *args);
66};
67
68struct arm_dma_buffer {
69 struct list_head list;
70 void *virt;
71 struct arm_dma_allocator *allocator;
72};
73
74static LIST_HEAD(arm_dma_bufs);
75static DEFINE_SPINLOCK(arm_dma_bufs_lock);
76
77static struct arm_dma_buffer *arm_dma_buffer_find(void *virt)
78{
79 struct arm_dma_buffer *buf, *found = NULL;
80 unsigned long flags;
81
82 spin_lock_irqsave(&arm_dma_bufs_lock, flags);
83 list_for_each_entry(buf, &arm_dma_bufs, list) {
84 if (buf->virt == virt) {
85 list_del(&buf->list);
86 found = buf;
87 break;
88 }
89 }
90 spin_unlock_irqrestore(&arm_dma_bufs_lock, flags);
91 return found;
92}
93
94/*
95 * The DMA API is built upon the notion of "buffer ownership". A buffer
96 * is either exclusively owned by the CPU (and therefore may be accessed
97 * by it) or exclusively owned by the DMA device. These helper functions
98 * represent the transitions between these two ownership states.
99 *
100 * Note, however, that on later ARMs, this notion does not work due to
101 * speculative prefetches. We model our approach on the assumption that
102 * the CPU does do speculative prefetches, which means we clean caches
103 * before transfers and delay cache invalidation until transfer completion.
104 *
105 */
106static void __dma_page_cpu_to_dev(struct page *, unsigned long,
107 size_t, enum dma_data_direction);
108static void __dma_page_dev_to_cpu(struct page *, unsigned long,
109 size_t, enum dma_data_direction);
110
111/**
112 * arm_dma_map_page - map a portion of a page for streaming DMA
113 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
114 * @page: page that buffer resides in
115 * @offset: offset into page for start of buffer
116 * @size: size of buffer to map
117 * @dir: DMA transfer direction
118 *
119 * Ensure that any data held in the cache is appropriately discarded
120 * or written back.
121 *
122 * The device owns this memory once this call has completed. The CPU
123 * can regain ownership by calling dma_unmap_page().
124 */
125static dma_addr_t arm_dma_map_page(struct device *dev, struct page *page,
126 unsigned long offset, size_t size, enum dma_data_direction dir,
127 unsigned long attrs)
128{
129 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
130 __dma_page_cpu_to_dev(page, offset, size, dir);
131 return pfn_to_dma(dev, page_to_pfn(page)) + offset;
132}
133
134static dma_addr_t arm_coherent_dma_map_page(struct device *dev, struct page *page,
135 unsigned long offset, size_t size, enum dma_data_direction dir,
136 unsigned long attrs)
137{
138 return pfn_to_dma(dev, page_to_pfn(page)) + offset;
139}
140
141/**
142 * arm_dma_unmap_page - unmap a buffer previously mapped through dma_map_page()
143 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
144 * @handle: DMA address of buffer
145 * @size: size of buffer (same as passed to dma_map_page)
146 * @dir: DMA transfer direction (same as passed to dma_map_page)
147 *
148 * Unmap a page streaming mode DMA translation. The handle and size
149 * must match what was provided in the previous dma_map_page() call.
150 * All other usages are undefined.
151 *
152 * After this call, reads by the CPU to the buffer are guaranteed to see
153 * whatever the device wrote there.
154 */
155static void arm_dma_unmap_page(struct device *dev, dma_addr_t handle,
156 size_t size, enum dma_data_direction dir, unsigned long attrs)
157{
158 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
159 __dma_page_dev_to_cpu(pfn_to_page(dma_to_pfn(dev, handle)),
160 handle & ~PAGE_MASK, size, dir);
161}
162
163static void arm_dma_sync_single_for_cpu(struct device *dev,
164 dma_addr_t handle, size_t size, enum dma_data_direction dir)
165{
166 unsigned int offset = handle & (PAGE_SIZE - 1);
167 struct page *page = pfn_to_page(dma_to_pfn(dev, handle-offset));
168 __dma_page_dev_to_cpu(page, offset, size, dir);
169}
170
171static void arm_dma_sync_single_for_device(struct device *dev,
172 dma_addr_t handle, size_t size, enum dma_data_direction dir)
173{
174 unsigned int offset = handle & (PAGE_SIZE - 1);
175 struct page *page = pfn_to_page(dma_to_pfn(dev, handle-offset));
176 __dma_page_cpu_to_dev(page, offset, size, dir);
177}
178
179/*
180 * Return whether the given device DMA address mask can be supported
181 * properly. For example, if your device can only drive the low 24-bits
182 * during bus mastering, then you would pass 0x00ffffff as the mask
183 * to this function.
184 */
185static int arm_dma_supported(struct device *dev, u64 mask)
186{
187 unsigned long max_dma_pfn = min(max_pfn - 1, arm_dma_pfn_limit);
188
189 /*
190 * Translate the device's DMA mask to a PFN limit. This
191 * PFN number includes the page which we can DMA to.
192 */
193 return dma_to_pfn(dev, mask) >= max_dma_pfn;
194}
195
196const struct dma_map_ops arm_dma_ops = {
197 .alloc = arm_dma_alloc,
198 .free = arm_dma_free,
199 .alloc_pages = dma_direct_alloc_pages,
200 .free_pages = dma_direct_free_pages,
201 .mmap = arm_dma_mmap,
202 .get_sgtable = arm_dma_get_sgtable,
203 .map_page = arm_dma_map_page,
204 .unmap_page = arm_dma_unmap_page,
205 .map_sg = arm_dma_map_sg,
206 .unmap_sg = arm_dma_unmap_sg,
207 .map_resource = dma_direct_map_resource,
208 .sync_single_for_cpu = arm_dma_sync_single_for_cpu,
209 .sync_single_for_device = arm_dma_sync_single_for_device,
210 .sync_sg_for_cpu = arm_dma_sync_sg_for_cpu,
211 .sync_sg_for_device = arm_dma_sync_sg_for_device,
212 .dma_supported = arm_dma_supported,
213 .get_required_mask = dma_direct_get_required_mask,
214};
215EXPORT_SYMBOL(arm_dma_ops);
216
217static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
218 dma_addr_t *handle, gfp_t gfp, unsigned long attrs);
219static void arm_coherent_dma_free(struct device *dev, size_t size, void *cpu_addr,
220 dma_addr_t handle, unsigned long attrs);
221static int arm_coherent_dma_mmap(struct device *dev, struct vm_area_struct *vma,
222 void *cpu_addr, dma_addr_t dma_addr, size_t size,
223 unsigned long attrs);
224
225const struct dma_map_ops arm_coherent_dma_ops = {
226 .alloc = arm_coherent_dma_alloc,
227 .free = arm_coherent_dma_free,
228 .alloc_pages = dma_direct_alloc_pages,
229 .free_pages = dma_direct_free_pages,
230 .mmap = arm_coherent_dma_mmap,
231 .get_sgtable = arm_dma_get_sgtable,
232 .map_page = arm_coherent_dma_map_page,
233 .map_sg = arm_dma_map_sg,
234 .map_resource = dma_direct_map_resource,
235 .dma_supported = arm_dma_supported,
236 .get_required_mask = dma_direct_get_required_mask,
237};
238EXPORT_SYMBOL(arm_coherent_dma_ops);
239
240static void __dma_clear_buffer(struct page *page, size_t size, int coherent_flag)
241{
242 /*
243 * Ensure that the allocated pages are zeroed, and that any data
244 * lurking in the kernel direct-mapped region is invalidated.
245 */
246 if (PageHighMem(page)) {
247 phys_addr_t base = __pfn_to_phys(page_to_pfn(page));
248 phys_addr_t end = base + size;
249 while (size > 0) {
250 void *ptr = kmap_atomic(page);
251 memset(ptr, 0, PAGE_SIZE);
252 if (coherent_flag != COHERENT)
253 dmac_flush_range(ptr, ptr + PAGE_SIZE);
254 kunmap_atomic(ptr);
255 page++;
256 size -= PAGE_SIZE;
257 }
258 if (coherent_flag != COHERENT)
259 outer_flush_range(base, end);
260 } else {
261 void *ptr = page_address(page);
262 memset(ptr, 0, size);
263 if (coherent_flag != COHERENT) {
264 dmac_flush_range(ptr, ptr + size);
265 outer_flush_range(__pa(ptr), __pa(ptr) + size);
266 }
267 }
268}
269
270/*
271 * Allocate a DMA buffer for 'dev' of size 'size' using the
272 * specified gfp mask. Note that 'size' must be page aligned.
273 */
274static struct page *__dma_alloc_buffer(struct device *dev, size_t size,
275 gfp_t gfp, int coherent_flag)
276{
277 unsigned long order = get_order(size);
278 struct page *page, *p, *e;
279
280 page = alloc_pages(gfp, order);
281 if (!page)
282 return NULL;
283
284 /*
285 * Now split the huge page and free the excess pages
286 */
287 split_page(page, order);
288 for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
289 __free_page(p);
290
291 __dma_clear_buffer(page, size, coherent_flag);
292
293 return page;
294}
295
296/*
297 * Free a DMA buffer. 'size' must be page aligned.
298 */
299static void __dma_free_buffer(struct page *page, size_t size)
300{
301 struct page *e = page + (size >> PAGE_SHIFT);
302
303 while (page < e) {
304 __free_page(page);
305 page++;
306 }
307}
308
309static void *__alloc_from_contiguous(struct device *dev, size_t size,
310 pgprot_t prot, struct page **ret_page,
311 const void *caller, bool want_vaddr,
312 int coherent_flag, gfp_t gfp);
313
314static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
315 pgprot_t prot, struct page **ret_page,
316 const void *caller, bool want_vaddr);
317
318#define DEFAULT_DMA_COHERENT_POOL_SIZE SZ_256K
319static struct gen_pool *atomic_pool __ro_after_init;
320
321static size_t atomic_pool_size __initdata = DEFAULT_DMA_COHERENT_POOL_SIZE;
322
323static int __init early_coherent_pool(char *p)
324{
325 atomic_pool_size = memparse(p, &p);
326 return 0;
327}
328early_param("coherent_pool", early_coherent_pool);
329
330/*
331 * Initialise the coherent pool for atomic allocations.
332 */
333static int __init atomic_pool_init(void)
334{
335 pgprot_t prot = pgprot_dmacoherent(PAGE_KERNEL);
336 gfp_t gfp = GFP_KERNEL | GFP_DMA;
337 struct page *page;
338 void *ptr;
339
340 atomic_pool = gen_pool_create(PAGE_SHIFT, -1);
341 if (!atomic_pool)
342 goto out;
343 /*
344 * The atomic pool is only used for non-coherent allocations
345 * so we must pass NORMAL for coherent_flag.
346 */
347 if (dev_get_cma_area(NULL))
348 ptr = __alloc_from_contiguous(NULL, atomic_pool_size, prot,
349 &page, atomic_pool_init, true, NORMAL,
350 GFP_KERNEL);
351 else
352 ptr = __alloc_remap_buffer(NULL, atomic_pool_size, gfp, prot,
353 &page, atomic_pool_init, true);
354 if (ptr) {
355 int ret;
356
357 ret = gen_pool_add_virt(atomic_pool, (unsigned long)ptr,
358 page_to_phys(page),
359 atomic_pool_size, -1);
360 if (ret)
361 goto destroy_genpool;
362
363 gen_pool_set_algo(atomic_pool,
364 gen_pool_first_fit_order_align,
365 NULL);
366 pr_info("DMA: preallocated %zu KiB pool for atomic coherent allocations\n",
367 atomic_pool_size / 1024);
368 return 0;
369 }
370
371destroy_genpool:
372 gen_pool_destroy(atomic_pool);
373 atomic_pool = NULL;
374out:
375 pr_err("DMA: failed to allocate %zu KiB pool for atomic coherent allocation\n",
376 atomic_pool_size / 1024);
377 return -ENOMEM;
378}
379/*
380 * CMA is activated by core_initcall, so we must be called after it.
381 */
382postcore_initcall(atomic_pool_init);
383
384struct dma_contig_early_reserve {
385 phys_addr_t base;
386 unsigned long size;
387};
388
389static struct dma_contig_early_reserve dma_mmu_remap[MAX_CMA_AREAS] __initdata;
390
391static int dma_mmu_remap_num __initdata;
392
393void __init dma_contiguous_early_fixup(phys_addr_t base, unsigned long size)
394{
395 dma_mmu_remap[dma_mmu_remap_num].base = base;
396 dma_mmu_remap[dma_mmu_remap_num].size = size;
397 dma_mmu_remap_num++;
398}
399
400void __init dma_contiguous_remap(void)
401{
402 int i;
403 for (i = 0; i < dma_mmu_remap_num; i++) {
404 phys_addr_t start = dma_mmu_remap[i].base;
405 phys_addr_t end = start + dma_mmu_remap[i].size;
406 struct map_desc map;
407 unsigned long addr;
408
409 if (end > arm_lowmem_limit)
410 end = arm_lowmem_limit;
411 if (start >= end)
412 continue;
413
414 map.pfn = __phys_to_pfn(start);
415 map.virtual = __phys_to_virt(start);
416 map.length = end - start;
417 map.type = MT_MEMORY_DMA_READY;
418
419 /*
420 * Clear previous low-memory mapping to ensure that the
421 * TLB does not see any conflicting entries, then flush
422 * the TLB of the old entries before creating new mappings.
423 *
424 * This ensures that any speculatively loaded TLB entries
425 * (even though they may be rare) can not cause any problems,
426 * and ensures that this code is architecturally compliant.
427 */
428 for (addr = __phys_to_virt(start); addr < __phys_to_virt(end);
429 addr += PMD_SIZE)
430 pmd_clear(pmd_off_k(addr));
431
432 flush_tlb_kernel_range(__phys_to_virt(start),
433 __phys_to_virt(end));
434
435 iotable_init(&map, 1);
436 }
437}
438
439static int __dma_update_pte(pte_t *pte, unsigned long addr, void *data)
440{
441 struct page *page = virt_to_page(addr);
442 pgprot_t prot = *(pgprot_t *)data;
443
444 set_pte_ext(pte, mk_pte(page, prot), 0);
445 return 0;
446}
447
448static void __dma_remap(struct page *page, size_t size, pgprot_t prot)
449{
450 unsigned long start = (unsigned long) page_address(page);
451 unsigned end = start + size;
452
453 apply_to_page_range(&init_mm, start, size, __dma_update_pte, &prot);
454 flush_tlb_kernel_range(start, end);
455}
456
457static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
458 pgprot_t prot, struct page **ret_page,
459 const void *caller, bool want_vaddr)
460{
461 struct page *page;
462 void *ptr = NULL;
463 /*
464 * __alloc_remap_buffer is only called when the device is
465 * non-coherent
466 */
467 page = __dma_alloc_buffer(dev, size, gfp, NORMAL);
468 if (!page)
469 return NULL;
470 if (!want_vaddr)
471 goto out;
472
473 ptr = dma_common_contiguous_remap(page, size, prot, caller);
474 if (!ptr) {
475 __dma_free_buffer(page, size);
476 return NULL;
477 }
478
479 out:
480 *ret_page = page;
481 return ptr;
482}
483
484static void *__alloc_from_pool(size_t size, struct page **ret_page)
485{
486 unsigned long val;
487 void *ptr = NULL;
488
489 if (!atomic_pool) {
490 WARN(1, "coherent pool not initialised!\n");
491 return NULL;
492 }
493
494 val = gen_pool_alloc(atomic_pool, size);
495 if (val) {
496 phys_addr_t phys = gen_pool_virt_to_phys(atomic_pool, val);
497
498 *ret_page = phys_to_page(phys);
499 ptr = (void *)val;
500 }
501
502 return ptr;
503}
504
505static bool __in_atomic_pool(void *start, size_t size)
506{
507 return gen_pool_has_addr(atomic_pool, (unsigned long)start, size);
508}
509
510static int __free_from_pool(void *start, size_t size)
511{
512 if (!__in_atomic_pool(start, size))
513 return 0;
514
515 gen_pool_free(atomic_pool, (unsigned long)start, size);
516
517 return 1;
518}
519
520static void *__alloc_from_contiguous(struct device *dev, size_t size,
521 pgprot_t prot, struct page **ret_page,
522 const void *caller, bool want_vaddr,
523 int coherent_flag, gfp_t gfp)
524{
525 unsigned long order = get_order(size);
526 size_t count = size >> PAGE_SHIFT;
527 struct page *page;
528 void *ptr = NULL;
529
530 page = dma_alloc_from_contiguous(dev, count, order, gfp & __GFP_NOWARN);
531 if (!page)
532 return NULL;
533
534 __dma_clear_buffer(page, size, coherent_flag);
535
536 if (!want_vaddr)
537 goto out;
538
539 if (PageHighMem(page)) {
540 ptr = dma_common_contiguous_remap(page, size, prot, caller);
541 if (!ptr) {
542 dma_release_from_contiguous(dev, page, count);
543 return NULL;
544 }
545 } else {
546 __dma_remap(page, size, prot);
547 ptr = page_address(page);
548 }
549
550 out:
551 *ret_page = page;
552 return ptr;
553}
554
555static void __free_from_contiguous(struct device *dev, struct page *page,
556 void *cpu_addr, size_t size, bool want_vaddr)
557{
558 if (want_vaddr) {
559 if (PageHighMem(page))
560 dma_common_free_remap(cpu_addr, size);
561 else
562 __dma_remap(page, size, PAGE_KERNEL);
563 }
564 dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT);
565}
566
567static inline pgprot_t __get_dma_pgprot(unsigned long attrs, pgprot_t prot)
568{
569 prot = (attrs & DMA_ATTR_WRITE_COMBINE) ?
570 pgprot_writecombine(prot) :
571 pgprot_dmacoherent(prot);
572 return prot;
573}
574
575static void *__alloc_simple_buffer(struct device *dev, size_t size, gfp_t gfp,
576 struct page **ret_page)
577{
578 struct page *page;
579 /* __alloc_simple_buffer is only called when the device is coherent */
580 page = __dma_alloc_buffer(dev, size, gfp, COHERENT);
581 if (!page)
582 return NULL;
583
584 *ret_page = page;
585 return page_address(page);
586}
587
588static void *simple_allocator_alloc(struct arm_dma_alloc_args *args,
589 struct page **ret_page)
590{
591 return __alloc_simple_buffer(args->dev, args->size, args->gfp,
592 ret_page);
593}
594
595static void simple_allocator_free(struct arm_dma_free_args *args)
596{
597 __dma_free_buffer(args->page, args->size);
598}
599
600static struct arm_dma_allocator simple_allocator = {
601 .alloc = simple_allocator_alloc,
602 .free = simple_allocator_free,
603};
604
605static void *cma_allocator_alloc(struct arm_dma_alloc_args *args,
606 struct page **ret_page)
607{
608 return __alloc_from_contiguous(args->dev, args->size, args->prot,
609 ret_page, args->caller,
610 args->want_vaddr, args->coherent_flag,
611 args->gfp);
612}
613
614static void cma_allocator_free(struct arm_dma_free_args *args)
615{
616 __free_from_contiguous(args->dev, args->page, args->cpu_addr,
617 args->size, args->want_vaddr);
618}
619
620static struct arm_dma_allocator cma_allocator = {
621 .alloc = cma_allocator_alloc,
622 .free = cma_allocator_free,
623};
624
625static void *pool_allocator_alloc(struct arm_dma_alloc_args *args,
626 struct page **ret_page)
627{
628 return __alloc_from_pool(args->size, ret_page);
629}
630
631static void pool_allocator_free(struct arm_dma_free_args *args)
632{
633 __free_from_pool(args->cpu_addr, args->size);
634}
635
636static struct arm_dma_allocator pool_allocator = {
637 .alloc = pool_allocator_alloc,
638 .free = pool_allocator_free,
639};
640
641static void *remap_allocator_alloc(struct arm_dma_alloc_args *args,
642 struct page **ret_page)
643{
644 return __alloc_remap_buffer(args->dev, args->size, args->gfp,
645 args->prot, ret_page, args->caller,
646 args->want_vaddr);
647}
648
649static void remap_allocator_free(struct arm_dma_free_args *args)
650{
651 if (args->want_vaddr)
652 dma_common_free_remap(args->cpu_addr, args->size);
653
654 __dma_free_buffer(args->page, args->size);
655}
656
657static struct arm_dma_allocator remap_allocator = {
658 .alloc = remap_allocator_alloc,
659 .free = remap_allocator_free,
660};
661
662static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
663 gfp_t gfp, pgprot_t prot, bool is_coherent,
664 unsigned long attrs, const void *caller)
665{
666 u64 mask = min_not_zero(dev->coherent_dma_mask, dev->bus_dma_limit);
667 struct page *page = NULL;
668 void *addr;
669 bool allowblock, cma;
670 struct arm_dma_buffer *buf;
671 struct arm_dma_alloc_args args = {
672 .dev = dev,
673 .size = PAGE_ALIGN(size),
674 .gfp = gfp,
675 .prot = prot,
676 .caller = caller,
677 .want_vaddr = ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) == 0),
678 .coherent_flag = is_coherent ? COHERENT : NORMAL,
679 };
680
681#ifdef CONFIG_DMA_API_DEBUG
682 u64 limit = (mask + 1) & ~mask;
683 if (limit && size >= limit) {
684 dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n",
685 size, mask);
686 return NULL;
687 }
688#endif
689
690 buf = kzalloc(sizeof(*buf),
691 gfp & ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM));
692 if (!buf)
693 return NULL;
694
695 if (mask < 0xffffffffULL)
696 gfp |= GFP_DMA;
697
698 /*
699 * Following is a work-around (a.k.a. hack) to prevent pages
700 * with __GFP_COMP being passed to split_page() which cannot
701 * handle them. The real problem is that this flag probably
702 * should be 0 on ARM as it is not supported on this
703 * platform; see CONFIG_HUGETLBFS.
704 */
705 gfp &= ~(__GFP_COMP);
706 args.gfp = gfp;
707
708 *handle = DMA_MAPPING_ERROR;
709 allowblock = gfpflags_allow_blocking(gfp);
710 cma = allowblock ? dev_get_cma_area(dev) : false;
711
712 if (cma)
713 buf->allocator = &cma_allocator;
714 else if (is_coherent)
715 buf->allocator = &simple_allocator;
716 else if (allowblock)
717 buf->allocator = &remap_allocator;
718 else
719 buf->allocator = &pool_allocator;
720
721 addr = buf->allocator->alloc(&args, &page);
722
723 if (page) {
724 unsigned long flags;
725
726 *handle = pfn_to_dma(dev, page_to_pfn(page));
727 buf->virt = args.want_vaddr ? addr : page;
728
729 spin_lock_irqsave(&arm_dma_bufs_lock, flags);
730 list_add(&buf->list, &arm_dma_bufs);
731 spin_unlock_irqrestore(&arm_dma_bufs_lock, flags);
732 } else {
733 kfree(buf);
734 }
735
736 return args.want_vaddr ? addr : page;
737}
738
739/*
740 * Allocate DMA-coherent memory space and return both the kernel remapped
741 * virtual and bus address for that space.
742 */
743void *arm_dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
744 gfp_t gfp, unsigned long attrs)
745{
746 pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
747
748 return __dma_alloc(dev, size, handle, gfp, prot, false,
749 attrs, __builtin_return_address(0));
750}
751
752static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
753 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
754{
755 return __dma_alloc(dev, size, handle, gfp, PAGE_KERNEL, true,
756 attrs, __builtin_return_address(0));
757}
758
759static int __arm_dma_mmap(struct device *dev, struct vm_area_struct *vma,
760 void *cpu_addr, dma_addr_t dma_addr, size_t size,
761 unsigned long attrs)
762{
763 int ret = -ENXIO;
764 unsigned long nr_vma_pages = vma_pages(vma);
765 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
766 unsigned long pfn = dma_to_pfn(dev, dma_addr);
767 unsigned long off = vma->vm_pgoff;
768
769 if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
770 return ret;
771
772 if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) {
773 ret = remap_pfn_range(vma, vma->vm_start,
774 pfn + off,
775 vma->vm_end - vma->vm_start,
776 vma->vm_page_prot);
777 }
778
779 return ret;
780}
781
782/*
783 * Create userspace mapping for the DMA-coherent memory.
784 */
785static int arm_coherent_dma_mmap(struct device *dev, struct vm_area_struct *vma,
786 void *cpu_addr, dma_addr_t dma_addr, size_t size,
787 unsigned long attrs)
788{
789 return __arm_dma_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
790}
791
792int arm_dma_mmap(struct device *dev, struct vm_area_struct *vma,
793 void *cpu_addr, dma_addr_t dma_addr, size_t size,
794 unsigned long attrs)
795{
796 vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot);
797 return __arm_dma_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
798}
799
800/*
801 * Free a buffer as defined by the above mapping.
802 */
803static void __arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
804 dma_addr_t handle, unsigned long attrs,
805 bool is_coherent)
806{
807 struct page *page = pfn_to_page(dma_to_pfn(dev, handle));
808 struct arm_dma_buffer *buf;
809 struct arm_dma_free_args args = {
810 .dev = dev,
811 .size = PAGE_ALIGN(size),
812 .cpu_addr = cpu_addr,
813 .page = page,
814 .want_vaddr = ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) == 0),
815 };
816
817 buf = arm_dma_buffer_find(cpu_addr);
818 if (WARN(!buf, "Freeing invalid buffer %p\n", cpu_addr))
819 return;
820
821 buf->allocator->free(&args);
822 kfree(buf);
823}
824
825void arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
826 dma_addr_t handle, unsigned long attrs)
827{
828 __arm_dma_free(dev, size, cpu_addr, handle, attrs, false);
829}
830
831static void arm_coherent_dma_free(struct device *dev, size_t size, void *cpu_addr,
832 dma_addr_t handle, unsigned long attrs)
833{
834 __arm_dma_free(dev, size, cpu_addr, handle, attrs, true);
835}
836
837int arm_dma_get_sgtable(struct device *dev, struct sg_table *sgt,
838 void *cpu_addr, dma_addr_t handle, size_t size,
839 unsigned long attrs)
840{
841 unsigned long pfn = dma_to_pfn(dev, handle);
842 struct page *page;
843 int ret;
844
845 /* If the PFN is not valid, we do not have a struct page */
846 if (!pfn_valid(pfn))
847 return -ENXIO;
848
849 page = pfn_to_page(pfn);
850
851 ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
852 if (unlikely(ret))
853 return ret;
854
855 sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
856 return 0;
857}
858
859static void dma_cache_maint_page(struct page *page, unsigned long offset,
860 size_t size, enum dma_data_direction dir,
861 void (*op)(const void *, size_t, int))
862{
863 unsigned long pfn;
864 size_t left = size;
865
866 pfn = page_to_pfn(page) + offset / PAGE_SIZE;
867 offset %= PAGE_SIZE;
868
869 /*
870 * A single sg entry may refer to multiple physically contiguous
871 * pages. But we still need to process highmem pages individually.
872 * If highmem is not configured then the bulk of this loop gets
873 * optimized out.
874 */
875 do {
876 size_t len = left;
877 void *vaddr;
878
879 page = pfn_to_page(pfn);
880
881 if (PageHighMem(page)) {
882 if (len + offset > PAGE_SIZE)
883 len = PAGE_SIZE - offset;
884
885 if (cache_is_vipt_nonaliasing()) {
886 vaddr = kmap_atomic(page);
887 op(vaddr + offset, len, dir);
888 kunmap_atomic(vaddr);
889 } else {
890 vaddr = kmap_high_get(page);
891 if (vaddr) {
892 op(vaddr + offset, len, dir);
893 kunmap_high(page);
894 }
895 }
896 } else {
897 vaddr = page_address(page) + offset;
898 op(vaddr, len, dir);
899 }
900 offset = 0;
901 pfn++;
902 left -= len;
903 } while (left);
904}
905
906/*
907 * Make an area consistent for devices.
908 * Note: Drivers should NOT use this function directly, as it will break
909 * platforms with CONFIG_DMABOUNCE.
910 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
911 */
912static void __dma_page_cpu_to_dev(struct page *page, unsigned long off,
913 size_t size, enum dma_data_direction dir)
914{
915 phys_addr_t paddr;
916
917 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
918
919 paddr = page_to_phys(page) + off;
920 if (dir == DMA_FROM_DEVICE) {
921 outer_inv_range(paddr, paddr + size);
922 } else {
923 outer_clean_range(paddr, paddr + size);
924 }
925 /* FIXME: non-speculating: flush on bidirectional mappings? */
926}
927
928static void __dma_page_dev_to_cpu(struct page *page, unsigned long off,
929 size_t size, enum dma_data_direction dir)
930{
931 phys_addr_t paddr = page_to_phys(page) + off;
932
933 /* FIXME: non-speculating: not required */
934 /* in any case, don't bother invalidating if DMA to device */
935 if (dir != DMA_TO_DEVICE) {
936 outer_inv_range(paddr, paddr + size);
937
938 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
939 }
940
941 /*
942 * Mark the D-cache clean for these pages to avoid extra flushing.
943 */
944 if (dir != DMA_TO_DEVICE && size >= PAGE_SIZE) {
945 unsigned long pfn;
946 size_t left = size;
947
948 pfn = page_to_pfn(page) + off / PAGE_SIZE;
949 off %= PAGE_SIZE;
950 if (off) {
951 pfn++;
952 left -= PAGE_SIZE - off;
953 }
954 while (left >= PAGE_SIZE) {
955 page = pfn_to_page(pfn++);
956 set_bit(PG_dcache_clean, &page->flags);
957 left -= PAGE_SIZE;
958 }
959 }
960}
961
962/**
963 * arm_dma_map_sg - map a set of SG buffers for streaming mode DMA
964 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
965 * @sg: list of buffers
966 * @nents: number of buffers to map
967 * @dir: DMA transfer direction
968 *
969 * Map a set of buffers described by scatterlist in streaming mode for DMA.
970 * This is the scatter-gather version of the dma_map_single interface.
971 * Here the scatter gather list elements are each tagged with the
972 * appropriate dma address and length. They are obtained via
973 * sg_dma_{address,length}.
974 *
975 * Device ownership issues as mentioned for dma_map_single are the same
976 * here.
977 */
978int arm_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
979 enum dma_data_direction dir, unsigned long attrs)
980{
981 const struct dma_map_ops *ops = get_dma_ops(dev);
982 struct scatterlist *s;
983 int i, j;
984
985 for_each_sg(sg, s, nents, i) {
986#ifdef CONFIG_NEED_SG_DMA_LENGTH
987 s->dma_length = s->length;
988#endif
989 s->dma_address = ops->map_page(dev, sg_page(s), s->offset,
990 s->length, dir, attrs);
991 if (dma_mapping_error(dev, s->dma_address))
992 goto bad_mapping;
993 }
994 return nents;
995
996 bad_mapping:
997 for_each_sg(sg, s, i, j)
998 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
999 return 0;
1000}
1001
1002/**
1003 * arm_dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
1004 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
1005 * @sg: list of buffers
1006 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
1007 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1008 *
1009 * Unmap a set of streaming mode DMA translations. Again, CPU access
1010 * rules concerning calls here are the same as for dma_unmap_single().
1011 */
1012void arm_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
1013 enum dma_data_direction dir, unsigned long attrs)
1014{
1015 const struct dma_map_ops *ops = get_dma_ops(dev);
1016 struct scatterlist *s;
1017
1018 int i;
1019
1020 for_each_sg(sg, s, nents, i)
1021 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
1022}
1023
1024/**
1025 * arm_dma_sync_sg_for_cpu
1026 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
1027 * @sg: list of buffers
1028 * @nents: number of buffers to map (returned from dma_map_sg)
1029 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1030 */
1031void arm_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1032 int nents, enum dma_data_direction dir)
1033{
1034 const struct dma_map_ops *ops = get_dma_ops(dev);
1035 struct scatterlist *s;
1036 int i;
1037
1038 for_each_sg(sg, s, nents, i)
1039 ops->sync_single_for_cpu(dev, sg_dma_address(s), s->length,
1040 dir);
1041}
1042
1043/**
1044 * arm_dma_sync_sg_for_device
1045 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
1046 * @sg: list of buffers
1047 * @nents: number of buffers to map (returned from dma_map_sg)
1048 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1049 */
1050void arm_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1051 int nents, enum dma_data_direction dir)
1052{
1053 const struct dma_map_ops *ops = get_dma_ops(dev);
1054 struct scatterlist *s;
1055 int i;
1056
1057 for_each_sg(sg, s, nents, i)
1058 ops->sync_single_for_device(dev, sg_dma_address(s), s->length,
1059 dir);
1060}
1061
1062static const struct dma_map_ops *arm_get_dma_map_ops(bool coherent)
1063{
1064 /*
1065 * When CONFIG_ARM_LPAE is set, physical address can extend above
1066 * 32-bits, which then can't be addressed by devices that only support
1067 * 32-bit DMA.
1068 * Use the generic dma-direct / swiotlb ops code in that case, as that
1069 * handles bounce buffering for us.
1070 */
1071 if (IS_ENABLED(CONFIG_ARM_LPAE))
1072 return NULL;
1073 return coherent ? &arm_coherent_dma_ops : &arm_dma_ops;
1074}
1075
1076#ifdef CONFIG_ARM_DMA_USE_IOMMU
1077
1078static int __dma_info_to_prot(enum dma_data_direction dir, unsigned long attrs)
1079{
1080 int prot = 0;
1081
1082 if (attrs & DMA_ATTR_PRIVILEGED)
1083 prot |= IOMMU_PRIV;
1084
1085 switch (dir) {
1086 case DMA_BIDIRECTIONAL:
1087 return prot | IOMMU_READ | IOMMU_WRITE;
1088 case DMA_TO_DEVICE:
1089 return prot | IOMMU_READ;
1090 case DMA_FROM_DEVICE:
1091 return prot | IOMMU_WRITE;
1092 default:
1093 return prot;
1094 }
1095}
1096
1097/* IOMMU */
1098
1099static int extend_iommu_mapping(struct dma_iommu_mapping *mapping);
1100
1101static inline dma_addr_t __alloc_iova(struct dma_iommu_mapping *mapping,
1102 size_t size)
1103{
1104 unsigned int order = get_order(size);
1105 unsigned int align = 0;
1106 unsigned int count, start;
1107 size_t mapping_size = mapping->bits << PAGE_SHIFT;
1108 unsigned long flags;
1109 dma_addr_t iova;
1110 int i;
1111
1112 if (order > CONFIG_ARM_DMA_IOMMU_ALIGNMENT)
1113 order = CONFIG_ARM_DMA_IOMMU_ALIGNMENT;
1114
1115 count = PAGE_ALIGN(size) >> PAGE_SHIFT;
1116 align = (1 << order) - 1;
1117
1118 spin_lock_irqsave(&mapping->lock, flags);
1119 for (i = 0; i < mapping->nr_bitmaps; i++) {
1120 start = bitmap_find_next_zero_area(mapping->bitmaps[i],
1121 mapping->bits, 0, count, align);
1122
1123 if (start > mapping->bits)
1124 continue;
1125
1126 bitmap_set(mapping->bitmaps[i], start, count);
1127 break;
1128 }
1129
1130 /*
1131 * No unused range found. Try to extend the existing mapping
1132 * and perform a second attempt to reserve an IO virtual
1133 * address range of size bytes.
1134 */
1135 if (i == mapping->nr_bitmaps) {
1136 if (extend_iommu_mapping(mapping)) {
1137 spin_unlock_irqrestore(&mapping->lock, flags);
1138 return DMA_MAPPING_ERROR;
1139 }
1140
1141 start = bitmap_find_next_zero_area(mapping->bitmaps[i],
1142 mapping->bits, 0, count, align);
1143
1144 if (start > mapping->bits) {
1145 spin_unlock_irqrestore(&mapping->lock, flags);
1146 return DMA_MAPPING_ERROR;
1147 }
1148
1149 bitmap_set(mapping->bitmaps[i], start, count);
1150 }
1151 spin_unlock_irqrestore(&mapping->lock, flags);
1152
1153 iova = mapping->base + (mapping_size * i);
1154 iova += start << PAGE_SHIFT;
1155
1156 return iova;
1157}
1158
1159static inline void __free_iova(struct dma_iommu_mapping *mapping,
1160 dma_addr_t addr, size_t size)
1161{
1162 unsigned int start, count;
1163 size_t mapping_size = mapping->bits << PAGE_SHIFT;
1164 unsigned long flags;
1165 dma_addr_t bitmap_base;
1166 u32 bitmap_index;
1167
1168 if (!size)
1169 return;
1170
1171 bitmap_index = (u32) (addr - mapping->base) / (u32) mapping_size;
1172 BUG_ON(addr < mapping->base || bitmap_index > mapping->extensions);
1173
1174 bitmap_base = mapping->base + mapping_size * bitmap_index;
1175
1176 start = (addr - bitmap_base) >> PAGE_SHIFT;
1177
1178 if (addr + size > bitmap_base + mapping_size) {
1179 /*
1180 * The address range to be freed reaches into the iova
1181 * range of the next bitmap. This should not happen as
1182 * we don't allow this in __alloc_iova (at the
1183 * moment).
1184 */
1185 BUG();
1186 } else
1187 count = size >> PAGE_SHIFT;
1188
1189 spin_lock_irqsave(&mapping->lock, flags);
1190 bitmap_clear(mapping->bitmaps[bitmap_index], start, count);
1191 spin_unlock_irqrestore(&mapping->lock, flags);
1192}
1193
1194/* We'll try 2M, 1M, 64K, and finally 4K; array must end with 0! */
1195static const int iommu_order_array[] = { 9, 8, 4, 0 };
1196
1197static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
1198 gfp_t gfp, unsigned long attrs,
1199 int coherent_flag)
1200{
1201 struct page **pages;
1202 int count = size >> PAGE_SHIFT;
1203 int array_size = count * sizeof(struct page *);
1204 int i = 0;
1205 int order_idx = 0;
1206
1207 if (array_size <= PAGE_SIZE)
1208 pages = kzalloc(array_size, GFP_KERNEL);
1209 else
1210 pages = vzalloc(array_size);
1211 if (!pages)
1212 return NULL;
1213
1214 if (attrs & DMA_ATTR_FORCE_CONTIGUOUS)
1215 {
1216 unsigned long order = get_order(size);
1217 struct page *page;
1218
1219 page = dma_alloc_from_contiguous(dev, count, order,
1220 gfp & __GFP_NOWARN);
1221 if (!page)
1222 goto error;
1223
1224 __dma_clear_buffer(page, size, coherent_flag);
1225
1226 for (i = 0; i < count; i++)
1227 pages[i] = page + i;
1228
1229 return pages;
1230 }
1231
1232 /* Go straight to 4K chunks if caller says it's OK. */
1233 if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)
1234 order_idx = ARRAY_SIZE(iommu_order_array) - 1;
1235
1236 /*
1237 * IOMMU can map any pages, so himem can also be used here
1238 */
1239 gfp |= __GFP_NOWARN | __GFP_HIGHMEM;
1240
1241 while (count) {
1242 int j, order;
1243
1244 order = iommu_order_array[order_idx];
1245
1246 /* Drop down when we get small */
1247 if (__fls(count) < order) {
1248 order_idx++;
1249 continue;
1250 }
1251
1252 if (order) {
1253 /* See if it's easy to allocate a high-order chunk */
1254 pages[i] = alloc_pages(gfp | __GFP_NORETRY, order);
1255
1256 /* Go down a notch at first sign of pressure */
1257 if (!pages[i]) {
1258 order_idx++;
1259 continue;
1260 }
1261 } else {
1262 pages[i] = alloc_pages(gfp, 0);
1263 if (!pages[i])
1264 goto error;
1265 }
1266
1267 if (order) {
1268 split_page(pages[i], order);
1269 j = 1 << order;
1270 while (--j)
1271 pages[i + j] = pages[i] + j;
1272 }
1273
1274 __dma_clear_buffer(pages[i], PAGE_SIZE << order, coherent_flag);
1275 i += 1 << order;
1276 count -= 1 << order;
1277 }
1278
1279 return pages;
1280error:
1281 while (i--)
1282 if (pages[i])
1283 __free_pages(pages[i], 0);
1284 kvfree(pages);
1285 return NULL;
1286}
1287
1288static int __iommu_free_buffer(struct device *dev, struct page **pages,
1289 size_t size, unsigned long attrs)
1290{
1291 int count = size >> PAGE_SHIFT;
1292 int i;
1293
1294 if (attrs & DMA_ATTR_FORCE_CONTIGUOUS) {
1295 dma_release_from_contiguous(dev, pages[0], count);
1296 } else {
1297 for (i = 0; i < count; i++)
1298 if (pages[i])
1299 __free_pages(pages[i], 0);
1300 }
1301
1302 kvfree(pages);
1303 return 0;
1304}
1305
1306/*
1307 * Create a mapping in device IO address space for specified pages
1308 */
1309static dma_addr_t
1310__iommu_create_mapping(struct device *dev, struct page **pages, size_t size,
1311 unsigned long attrs)
1312{
1313 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1314 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
1315 dma_addr_t dma_addr, iova;
1316 int i;
1317
1318 dma_addr = __alloc_iova(mapping, size);
1319 if (dma_addr == DMA_MAPPING_ERROR)
1320 return dma_addr;
1321
1322 iova = dma_addr;
1323 for (i = 0; i < count; ) {
1324 int ret;
1325
1326 unsigned int next_pfn = page_to_pfn(pages[i]) + 1;
1327 phys_addr_t phys = page_to_phys(pages[i]);
1328 unsigned int len, j;
1329
1330 for (j = i + 1; j < count; j++, next_pfn++)
1331 if (page_to_pfn(pages[j]) != next_pfn)
1332 break;
1333
1334 len = (j - i) << PAGE_SHIFT;
1335 ret = iommu_map(mapping->domain, iova, phys, len,
1336 __dma_info_to_prot(DMA_BIDIRECTIONAL, attrs));
1337 if (ret < 0)
1338 goto fail;
1339 iova += len;
1340 i = j;
1341 }
1342 return dma_addr;
1343fail:
1344 iommu_unmap(mapping->domain, dma_addr, iova-dma_addr);
1345 __free_iova(mapping, dma_addr, size);
1346 return DMA_MAPPING_ERROR;
1347}
1348
1349static int __iommu_remove_mapping(struct device *dev, dma_addr_t iova, size_t size)
1350{
1351 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1352
1353 /*
1354 * add optional in-page offset from iova to size and align
1355 * result to page size
1356 */
1357 size = PAGE_ALIGN((iova & ~PAGE_MASK) + size);
1358 iova &= PAGE_MASK;
1359
1360 iommu_unmap(mapping->domain, iova, size);
1361 __free_iova(mapping, iova, size);
1362 return 0;
1363}
1364
1365static struct page **__atomic_get_pages(void *addr)
1366{
1367 struct page *page;
1368 phys_addr_t phys;
1369
1370 phys = gen_pool_virt_to_phys(atomic_pool, (unsigned long)addr);
1371 page = phys_to_page(phys);
1372
1373 return (struct page **)page;
1374}
1375
1376static struct page **__iommu_get_pages(void *cpu_addr, unsigned long attrs)
1377{
1378 if (__in_atomic_pool(cpu_addr, PAGE_SIZE))
1379 return __atomic_get_pages(cpu_addr);
1380
1381 if (attrs & DMA_ATTR_NO_KERNEL_MAPPING)
1382 return cpu_addr;
1383
1384 return dma_common_find_pages(cpu_addr);
1385}
1386
1387static void *__iommu_alloc_simple(struct device *dev, size_t size, gfp_t gfp,
1388 dma_addr_t *handle, int coherent_flag,
1389 unsigned long attrs)
1390{
1391 struct page *page;
1392 void *addr;
1393
1394 if (coherent_flag == COHERENT)
1395 addr = __alloc_simple_buffer(dev, size, gfp, &page);
1396 else
1397 addr = __alloc_from_pool(size, &page);
1398 if (!addr)
1399 return NULL;
1400
1401 *handle = __iommu_create_mapping(dev, &page, size, attrs);
1402 if (*handle == DMA_MAPPING_ERROR)
1403 goto err_mapping;
1404
1405 return addr;
1406
1407err_mapping:
1408 __free_from_pool(addr, size);
1409 return NULL;
1410}
1411
1412static void __iommu_free_atomic(struct device *dev, void *cpu_addr,
1413 dma_addr_t handle, size_t size, int coherent_flag)
1414{
1415 __iommu_remove_mapping(dev, handle, size);
1416 if (coherent_flag == COHERENT)
1417 __dma_free_buffer(virt_to_page(cpu_addr), size);
1418 else
1419 __free_from_pool(cpu_addr, size);
1420}
1421
1422static void *__arm_iommu_alloc_attrs(struct device *dev, size_t size,
1423 dma_addr_t *handle, gfp_t gfp, unsigned long attrs,
1424 int coherent_flag)
1425{
1426 pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
1427 struct page **pages;
1428 void *addr = NULL;
1429
1430 *handle = DMA_MAPPING_ERROR;
1431 size = PAGE_ALIGN(size);
1432
1433 if (coherent_flag == COHERENT || !gfpflags_allow_blocking(gfp))
1434 return __iommu_alloc_simple(dev, size, gfp, handle,
1435 coherent_flag, attrs);
1436
1437 /*
1438 * Following is a work-around (a.k.a. hack) to prevent pages
1439 * with __GFP_COMP being passed to split_page() which cannot
1440 * handle them. The real problem is that this flag probably
1441 * should be 0 on ARM as it is not supported on this
1442 * platform; see CONFIG_HUGETLBFS.
1443 */
1444 gfp &= ~(__GFP_COMP);
1445
1446 pages = __iommu_alloc_buffer(dev, size, gfp, attrs, coherent_flag);
1447 if (!pages)
1448 return NULL;
1449
1450 *handle = __iommu_create_mapping(dev, pages, size, attrs);
1451 if (*handle == DMA_MAPPING_ERROR)
1452 goto err_buffer;
1453
1454 if (attrs & DMA_ATTR_NO_KERNEL_MAPPING)
1455 return pages;
1456
1457 addr = dma_common_pages_remap(pages, size, prot,
1458 __builtin_return_address(0));
1459 if (!addr)
1460 goto err_mapping;
1461
1462 return addr;
1463
1464err_mapping:
1465 __iommu_remove_mapping(dev, *handle, size);
1466err_buffer:
1467 __iommu_free_buffer(dev, pages, size, attrs);
1468 return NULL;
1469}
1470
1471static void *arm_iommu_alloc_attrs(struct device *dev, size_t size,
1472 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1473{
1474 return __arm_iommu_alloc_attrs(dev, size, handle, gfp, attrs, NORMAL);
1475}
1476
1477static void *arm_coherent_iommu_alloc_attrs(struct device *dev, size_t size,
1478 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1479{
1480 return __arm_iommu_alloc_attrs(dev, size, handle, gfp, attrs, COHERENT);
1481}
1482
1483static int __arm_iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
1484 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1485 unsigned long attrs)
1486{
1487 struct page **pages = __iommu_get_pages(cpu_addr, attrs);
1488 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
1489 int err;
1490
1491 if (!pages)
1492 return -ENXIO;
1493
1494 if (vma->vm_pgoff >= nr_pages)
1495 return -ENXIO;
1496
1497 err = vm_map_pages(vma, pages, nr_pages);
1498 if (err)
1499 pr_err("Remapping memory failed: %d\n", err);
1500
1501 return err;
1502}
1503static int arm_iommu_mmap_attrs(struct device *dev,
1504 struct vm_area_struct *vma, void *cpu_addr,
1505 dma_addr_t dma_addr, size_t size, unsigned long attrs)
1506{
1507 vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot);
1508
1509 return __arm_iommu_mmap_attrs(dev, vma, cpu_addr, dma_addr, size, attrs);
1510}
1511
1512static int arm_coherent_iommu_mmap_attrs(struct device *dev,
1513 struct vm_area_struct *vma, void *cpu_addr,
1514 dma_addr_t dma_addr, size_t size, unsigned long attrs)
1515{
1516 return __arm_iommu_mmap_attrs(dev, vma, cpu_addr, dma_addr, size, attrs);
1517}
1518
1519/*
1520 * free a page as defined by the above mapping.
1521 * Must not be called with IRQs disabled.
1522 */
1523static void __arm_iommu_free_attrs(struct device *dev, size_t size, void *cpu_addr,
1524 dma_addr_t handle, unsigned long attrs, int coherent_flag)
1525{
1526 struct page **pages;
1527 size = PAGE_ALIGN(size);
1528
1529 if (coherent_flag == COHERENT || __in_atomic_pool(cpu_addr, size)) {
1530 __iommu_free_atomic(dev, cpu_addr, handle, size, coherent_flag);
1531 return;
1532 }
1533
1534 pages = __iommu_get_pages(cpu_addr, attrs);
1535 if (!pages) {
1536 WARN(1, "trying to free invalid coherent area: %p\n", cpu_addr);
1537 return;
1538 }
1539
1540 if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) == 0)
1541 dma_common_free_remap(cpu_addr, size);
1542
1543 __iommu_remove_mapping(dev, handle, size);
1544 __iommu_free_buffer(dev, pages, size, attrs);
1545}
1546
1547static void arm_iommu_free_attrs(struct device *dev, size_t size,
1548 void *cpu_addr, dma_addr_t handle,
1549 unsigned long attrs)
1550{
1551 __arm_iommu_free_attrs(dev, size, cpu_addr, handle, attrs, NORMAL);
1552}
1553
1554static void arm_coherent_iommu_free_attrs(struct device *dev, size_t size,
1555 void *cpu_addr, dma_addr_t handle, unsigned long attrs)
1556{
1557 __arm_iommu_free_attrs(dev, size, cpu_addr, handle, attrs, COHERENT);
1558}
1559
1560static int arm_iommu_get_sgtable(struct device *dev, struct sg_table *sgt,
1561 void *cpu_addr, dma_addr_t dma_addr,
1562 size_t size, unsigned long attrs)
1563{
1564 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
1565 struct page **pages = __iommu_get_pages(cpu_addr, attrs);
1566
1567 if (!pages)
1568 return -ENXIO;
1569
1570 return sg_alloc_table_from_pages(sgt, pages, count, 0, size,
1571 GFP_KERNEL);
1572}
1573
1574/*
1575 * Map a part of the scatter-gather list into contiguous io address space
1576 */
1577static int __map_sg_chunk(struct device *dev, struct scatterlist *sg,
1578 size_t size, dma_addr_t *handle,
1579 enum dma_data_direction dir, unsigned long attrs,
1580 bool is_coherent)
1581{
1582 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1583 dma_addr_t iova, iova_base;
1584 int ret = 0;
1585 unsigned int count;
1586 struct scatterlist *s;
1587 int prot;
1588
1589 size = PAGE_ALIGN(size);
1590 *handle = DMA_MAPPING_ERROR;
1591
1592 iova_base = iova = __alloc_iova(mapping, size);
1593 if (iova == DMA_MAPPING_ERROR)
1594 return -ENOMEM;
1595
1596 for (count = 0, s = sg; count < (size >> PAGE_SHIFT); s = sg_next(s)) {
1597 phys_addr_t phys = page_to_phys(sg_page(s));
1598 unsigned int len = PAGE_ALIGN(s->offset + s->length);
1599
1600 if (!is_coherent && (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1601 __dma_page_cpu_to_dev(sg_page(s), s->offset, s->length, dir);
1602
1603 prot = __dma_info_to_prot(dir, attrs);
1604
1605 ret = iommu_map(mapping->domain, iova, phys, len, prot);
1606 if (ret < 0)
1607 goto fail;
1608 count += len >> PAGE_SHIFT;
1609 iova += len;
1610 }
1611 *handle = iova_base;
1612
1613 return 0;
1614fail:
1615 iommu_unmap(mapping->domain, iova_base, count * PAGE_SIZE);
1616 __free_iova(mapping, iova_base, size);
1617 return ret;
1618}
1619
1620static int __iommu_map_sg(struct device *dev, struct scatterlist *sg, int nents,
1621 enum dma_data_direction dir, unsigned long attrs,
1622 bool is_coherent)
1623{
1624 struct scatterlist *s = sg, *dma = sg, *start = sg;
1625 int i, count = 0;
1626 unsigned int offset = s->offset;
1627 unsigned int size = s->offset + s->length;
1628 unsigned int max = dma_get_max_seg_size(dev);
1629
1630 for (i = 1; i < nents; i++) {
1631 s = sg_next(s);
1632
1633 s->dma_address = DMA_MAPPING_ERROR;
1634 s->dma_length = 0;
1635
1636 if (s->offset || (size & ~PAGE_MASK) || size + s->length > max) {
1637 if (__map_sg_chunk(dev, start, size, &dma->dma_address,
1638 dir, attrs, is_coherent) < 0)
1639 goto bad_mapping;
1640
1641 dma->dma_address += offset;
1642 dma->dma_length = size - offset;
1643
1644 size = offset = s->offset;
1645 start = s;
1646 dma = sg_next(dma);
1647 count += 1;
1648 }
1649 size += s->length;
1650 }
1651 if (__map_sg_chunk(dev, start, size, &dma->dma_address, dir, attrs,
1652 is_coherent) < 0)
1653 goto bad_mapping;
1654
1655 dma->dma_address += offset;
1656 dma->dma_length = size - offset;
1657
1658 return count+1;
1659
1660bad_mapping:
1661 for_each_sg(sg, s, count, i)
1662 __iommu_remove_mapping(dev, sg_dma_address(s), sg_dma_len(s));
1663 return 0;
1664}
1665
1666/**
1667 * arm_coherent_iommu_map_sg - map a set of SG buffers for streaming mode DMA
1668 * @dev: valid struct device pointer
1669 * @sg: list of buffers
1670 * @nents: number of buffers to map
1671 * @dir: DMA transfer direction
1672 *
1673 * Map a set of i/o coherent buffers described by scatterlist in streaming
1674 * mode for DMA. The scatter gather list elements are merged together (if
1675 * possible) and tagged with the appropriate dma address and length. They are
1676 * obtained via sg_dma_{address,length}.
1677 */
1678static int arm_coherent_iommu_map_sg(struct device *dev, struct scatterlist *sg,
1679 int nents, enum dma_data_direction dir, unsigned long attrs)
1680{
1681 return __iommu_map_sg(dev, sg, nents, dir, attrs, true);
1682}
1683
1684/**
1685 * arm_iommu_map_sg - map a set of SG buffers for streaming mode DMA
1686 * @dev: valid struct device pointer
1687 * @sg: list of buffers
1688 * @nents: number of buffers to map
1689 * @dir: DMA transfer direction
1690 *
1691 * Map a set of buffers described by scatterlist in streaming mode for DMA.
1692 * The scatter gather list elements are merged together (if possible) and
1693 * tagged with the appropriate dma address and length. They are obtained via
1694 * sg_dma_{address,length}.
1695 */
1696static int arm_iommu_map_sg(struct device *dev, struct scatterlist *sg,
1697 int nents, enum dma_data_direction dir, unsigned long attrs)
1698{
1699 return __iommu_map_sg(dev, sg, nents, dir, attrs, false);
1700}
1701
1702static void __iommu_unmap_sg(struct device *dev, struct scatterlist *sg,
1703 int nents, enum dma_data_direction dir,
1704 unsigned long attrs, bool is_coherent)
1705{
1706 struct scatterlist *s;
1707 int i;
1708
1709 for_each_sg(sg, s, nents, i) {
1710 if (sg_dma_len(s))
1711 __iommu_remove_mapping(dev, sg_dma_address(s),
1712 sg_dma_len(s));
1713 if (!is_coherent && (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1714 __dma_page_dev_to_cpu(sg_page(s), s->offset,
1715 s->length, dir);
1716 }
1717}
1718
1719/**
1720 * arm_coherent_iommu_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
1721 * @dev: valid struct device pointer
1722 * @sg: list of buffers
1723 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
1724 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1725 *
1726 * Unmap a set of streaming mode DMA translations. Again, CPU access
1727 * rules concerning calls here are the same as for dma_unmap_single().
1728 */
1729static void arm_coherent_iommu_unmap_sg(struct device *dev,
1730 struct scatterlist *sg, int nents, enum dma_data_direction dir,
1731 unsigned long attrs)
1732{
1733 __iommu_unmap_sg(dev, sg, nents, dir, attrs, true);
1734}
1735
1736/**
1737 * arm_iommu_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
1738 * @dev: valid struct device pointer
1739 * @sg: list of buffers
1740 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
1741 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1742 *
1743 * Unmap a set of streaming mode DMA translations. Again, CPU access
1744 * rules concerning calls here are the same as for dma_unmap_single().
1745 */
1746static void arm_iommu_unmap_sg(struct device *dev,
1747 struct scatterlist *sg, int nents,
1748 enum dma_data_direction dir,
1749 unsigned long attrs)
1750{
1751 __iommu_unmap_sg(dev, sg, nents, dir, attrs, false);
1752}
1753
1754/**
1755 * arm_iommu_sync_sg_for_cpu
1756 * @dev: valid struct device pointer
1757 * @sg: list of buffers
1758 * @nents: number of buffers to map (returned from dma_map_sg)
1759 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1760 */
1761static void arm_iommu_sync_sg_for_cpu(struct device *dev,
1762 struct scatterlist *sg,
1763 int nents, enum dma_data_direction dir)
1764{
1765 struct scatterlist *s;
1766 int i;
1767
1768 for_each_sg(sg, s, nents, i)
1769 __dma_page_dev_to_cpu(sg_page(s), s->offset, s->length, dir);
1770
1771}
1772
1773/**
1774 * arm_iommu_sync_sg_for_device
1775 * @dev: valid struct device pointer
1776 * @sg: list of buffers
1777 * @nents: number of buffers to map (returned from dma_map_sg)
1778 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1779 */
1780static void arm_iommu_sync_sg_for_device(struct device *dev,
1781 struct scatterlist *sg,
1782 int nents, enum dma_data_direction dir)
1783{
1784 struct scatterlist *s;
1785 int i;
1786
1787 for_each_sg(sg, s, nents, i)
1788 __dma_page_cpu_to_dev(sg_page(s), s->offset, s->length, dir);
1789}
1790
1791
1792/**
1793 * arm_coherent_iommu_map_page
1794 * @dev: valid struct device pointer
1795 * @page: page that buffer resides in
1796 * @offset: offset into page for start of buffer
1797 * @size: size of buffer to map
1798 * @dir: DMA transfer direction
1799 *
1800 * Coherent IOMMU aware version of arm_dma_map_page()
1801 */
1802static dma_addr_t arm_coherent_iommu_map_page(struct device *dev, struct page *page,
1803 unsigned long offset, size_t size, enum dma_data_direction dir,
1804 unsigned long attrs)
1805{
1806 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1807 dma_addr_t dma_addr;
1808 int ret, prot, len = PAGE_ALIGN(size + offset);
1809
1810 dma_addr = __alloc_iova(mapping, len);
1811 if (dma_addr == DMA_MAPPING_ERROR)
1812 return dma_addr;
1813
1814 prot = __dma_info_to_prot(dir, attrs);
1815
1816 ret = iommu_map(mapping->domain, dma_addr, page_to_phys(page), len, prot);
1817 if (ret < 0)
1818 goto fail;
1819
1820 return dma_addr + offset;
1821fail:
1822 __free_iova(mapping, dma_addr, len);
1823 return DMA_MAPPING_ERROR;
1824}
1825
1826/**
1827 * arm_iommu_map_page
1828 * @dev: valid struct device pointer
1829 * @page: page that buffer resides in
1830 * @offset: offset into page for start of buffer
1831 * @size: size of buffer to map
1832 * @dir: DMA transfer direction
1833 *
1834 * IOMMU aware version of arm_dma_map_page()
1835 */
1836static dma_addr_t arm_iommu_map_page(struct device *dev, struct page *page,
1837 unsigned long offset, size_t size, enum dma_data_direction dir,
1838 unsigned long attrs)
1839{
1840 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1841 __dma_page_cpu_to_dev(page, offset, size, dir);
1842
1843 return arm_coherent_iommu_map_page(dev, page, offset, size, dir, attrs);
1844}
1845
1846/**
1847 * arm_coherent_iommu_unmap_page
1848 * @dev: valid struct device pointer
1849 * @handle: DMA address of buffer
1850 * @size: size of buffer (same as passed to dma_map_page)
1851 * @dir: DMA transfer direction (same as passed to dma_map_page)
1852 *
1853 * Coherent IOMMU aware version of arm_dma_unmap_page()
1854 */
1855static void arm_coherent_iommu_unmap_page(struct device *dev, dma_addr_t handle,
1856 size_t size, enum dma_data_direction dir, unsigned long attrs)
1857{
1858 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1859 dma_addr_t iova = handle & PAGE_MASK;
1860 int offset = handle & ~PAGE_MASK;
1861 int len = PAGE_ALIGN(size + offset);
1862
1863 if (!iova)
1864 return;
1865
1866 iommu_unmap(mapping->domain, iova, len);
1867 __free_iova(mapping, iova, len);
1868}
1869
1870/**
1871 * arm_iommu_unmap_page
1872 * @dev: valid struct device pointer
1873 * @handle: DMA address of buffer
1874 * @size: size of buffer (same as passed to dma_map_page)
1875 * @dir: DMA transfer direction (same as passed to dma_map_page)
1876 *
1877 * IOMMU aware version of arm_dma_unmap_page()
1878 */
1879static void arm_iommu_unmap_page(struct device *dev, dma_addr_t handle,
1880 size_t size, enum dma_data_direction dir, unsigned long attrs)
1881{
1882 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1883 dma_addr_t iova = handle & PAGE_MASK;
1884 struct page *page = phys_to_page(iommu_iova_to_phys(mapping->domain, iova));
1885 int offset = handle & ~PAGE_MASK;
1886 int len = PAGE_ALIGN(size + offset);
1887
1888 if (!iova)
1889 return;
1890
1891 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1892 __dma_page_dev_to_cpu(page, offset, size, dir);
1893
1894 iommu_unmap(mapping->domain, iova, len);
1895 __free_iova(mapping, iova, len);
1896}
1897
1898/**
1899 * arm_iommu_map_resource - map a device resource for DMA
1900 * @dev: valid struct device pointer
1901 * @phys_addr: physical address of resource
1902 * @size: size of resource to map
1903 * @dir: DMA transfer direction
1904 */
1905static dma_addr_t arm_iommu_map_resource(struct device *dev,
1906 phys_addr_t phys_addr, size_t size,
1907 enum dma_data_direction dir, unsigned long attrs)
1908{
1909 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1910 dma_addr_t dma_addr;
1911 int ret, prot;
1912 phys_addr_t addr = phys_addr & PAGE_MASK;
1913 unsigned int offset = phys_addr & ~PAGE_MASK;
1914 size_t len = PAGE_ALIGN(size + offset);
1915
1916 dma_addr = __alloc_iova(mapping, len);
1917 if (dma_addr == DMA_MAPPING_ERROR)
1918 return dma_addr;
1919
1920 prot = __dma_info_to_prot(dir, attrs) | IOMMU_MMIO;
1921
1922 ret = iommu_map(mapping->domain, dma_addr, addr, len, prot);
1923 if (ret < 0)
1924 goto fail;
1925
1926 return dma_addr + offset;
1927fail:
1928 __free_iova(mapping, dma_addr, len);
1929 return DMA_MAPPING_ERROR;
1930}
1931
1932/**
1933 * arm_iommu_unmap_resource - unmap a device DMA resource
1934 * @dev: valid struct device pointer
1935 * @dma_handle: DMA address to resource
1936 * @size: size of resource to map
1937 * @dir: DMA transfer direction
1938 */
1939static void arm_iommu_unmap_resource(struct device *dev, dma_addr_t dma_handle,
1940 size_t size, enum dma_data_direction dir,
1941 unsigned long attrs)
1942{
1943 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1944 dma_addr_t iova = dma_handle & PAGE_MASK;
1945 unsigned int offset = dma_handle & ~PAGE_MASK;
1946 size_t len = PAGE_ALIGN(size + offset);
1947
1948 if (!iova)
1949 return;
1950
1951 iommu_unmap(mapping->domain, iova, len);
1952 __free_iova(mapping, iova, len);
1953}
1954
1955static void arm_iommu_sync_single_for_cpu(struct device *dev,
1956 dma_addr_t handle, size_t size, enum dma_data_direction dir)
1957{
1958 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1959 dma_addr_t iova = handle & PAGE_MASK;
1960 struct page *page = phys_to_page(iommu_iova_to_phys(mapping->domain, iova));
1961 unsigned int offset = handle & ~PAGE_MASK;
1962
1963 if (!iova)
1964 return;
1965
1966 __dma_page_dev_to_cpu(page, offset, size, dir);
1967}
1968
1969static void arm_iommu_sync_single_for_device(struct device *dev,
1970 dma_addr_t handle, size_t size, enum dma_data_direction dir)
1971{
1972 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1973 dma_addr_t iova = handle & PAGE_MASK;
1974 struct page *page = phys_to_page(iommu_iova_to_phys(mapping->domain, iova));
1975 unsigned int offset = handle & ~PAGE_MASK;
1976
1977 if (!iova)
1978 return;
1979
1980 __dma_page_cpu_to_dev(page, offset, size, dir);
1981}
1982
1983static const struct dma_map_ops iommu_ops = {
1984 .alloc = arm_iommu_alloc_attrs,
1985 .free = arm_iommu_free_attrs,
1986 .mmap = arm_iommu_mmap_attrs,
1987 .get_sgtable = arm_iommu_get_sgtable,
1988
1989 .map_page = arm_iommu_map_page,
1990 .unmap_page = arm_iommu_unmap_page,
1991 .sync_single_for_cpu = arm_iommu_sync_single_for_cpu,
1992 .sync_single_for_device = arm_iommu_sync_single_for_device,
1993
1994 .map_sg = arm_iommu_map_sg,
1995 .unmap_sg = arm_iommu_unmap_sg,
1996 .sync_sg_for_cpu = arm_iommu_sync_sg_for_cpu,
1997 .sync_sg_for_device = arm_iommu_sync_sg_for_device,
1998
1999 .map_resource = arm_iommu_map_resource,
2000 .unmap_resource = arm_iommu_unmap_resource,
2001
2002 .dma_supported = arm_dma_supported,
2003};
2004
2005static const struct dma_map_ops iommu_coherent_ops = {
2006 .alloc = arm_coherent_iommu_alloc_attrs,
2007 .free = arm_coherent_iommu_free_attrs,
2008 .mmap = arm_coherent_iommu_mmap_attrs,
2009 .get_sgtable = arm_iommu_get_sgtable,
2010
2011 .map_page = arm_coherent_iommu_map_page,
2012 .unmap_page = arm_coherent_iommu_unmap_page,
2013
2014 .map_sg = arm_coherent_iommu_map_sg,
2015 .unmap_sg = arm_coherent_iommu_unmap_sg,
2016
2017 .map_resource = arm_iommu_map_resource,
2018 .unmap_resource = arm_iommu_unmap_resource,
2019
2020 .dma_supported = arm_dma_supported,
2021};
2022
2023/**
2024 * arm_iommu_create_mapping
2025 * @bus: pointer to the bus holding the client device (for IOMMU calls)
2026 * @base: start address of the valid IO address space
2027 * @size: maximum size of the valid IO address space
2028 *
2029 * Creates a mapping structure which holds information about used/unused
2030 * IO address ranges, which is required to perform memory allocation and
2031 * mapping with IOMMU aware functions.
2032 *
2033 * The client device need to be attached to the mapping with
2034 * arm_iommu_attach_device function.
2035 */
2036struct dma_iommu_mapping *
2037arm_iommu_create_mapping(struct bus_type *bus, dma_addr_t base, u64 size)
2038{
2039 unsigned int bits = size >> PAGE_SHIFT;
2040 unsigned int bitmap_size = BITS_TO_LONGS(bits) * sizeof(long);
2041 struct dma_iommu_mapping *mapping;
2042 int extensions = 1;
2043 int err = -ENOMEM;
2044
2045 /* currently only 32-bit DMA address space is supported */
2046 if (size > DMA_BIT_MASK(32) + 1)
2047 return ERR_PTR(-ERANGE);
2048
2049 if (!bitmap_size)
2050 return ERR_PTR(-EINVAL);
2051
2052 if (bitmap_size > PAGE_SIZE) {
2053 extensions = bitmap_size / PAGE_SIZE;
2054 bitmap_size = PAGE_SIZE;
2055 }
2056
2057 mapping = kzalloc(sizeof(struct dma_iommu_mapping), GFP_KERNEL);
2058 if (!mapping)
2059 goto err;
2060
2061 mapping->bitmap_size = bitmap_size;
2062 mapping->bitmaps = kcalloc(extensions, sizeof(unsigned long *),
2063 GFP_KERNEL);
2064 if (!mapping->bitmaps)
2065 goto err2;
2066
2067 mapping->bitmaps[0] = kzalloc(bitmap_size, GFP_KERNEL);
2068 if (!mapping->bitmaps[0])
2069 goto err3;
2070
2071 mapping->nr_bitmaps = 1;
2072 mapping->extensions = extensions;
2073 mapping->base = base;
2074 mapping->bits = BITS_PER_BYTE * bitmap_size;
2075
2076 spin_lock_init(&mapping->lock);
2077
2078 mapping->domain = iommu_domain_alloc(bus);
2079 if (!mapping->domain)
2080 goto err4;
2081
2082 kref_init(&mapping->kref);
2083 return mapping;
2084err4:
2085 kfree(mapping->bitmaps[0]);
2086err3:
2087 kfree(mapping->bitmaps);
2088err2:
2089 kfree(mapping);
2090err:
2091 return ERR_PTR(err);
2092}
2093EXPORT_SYMBOL_GPL(arm_iommu_create_mapping);
2094
2095static void release_iommu_mapping(struct kref *kref)
2096{
2097 int i;
2098 struct dma_iommu_mapping *mapping =
2099 container_of(kref, struct dma_iommu_mapping, kref);
2100
2101 iommu_domain_free(mapping->domain);
2102 for (i = 0; i < mapping->nr_bitmaps; i++)
2103 kfree(mapping->bitmaps[i]);
2104 kfree(mapping->bitmaps);
2105 kfree(mapping);
2106}
2107
2108static int extend_iommu_mapping(struct dma_iommu_mapping *mapping)
2109{
2110 int next_bitmap;
2111
2112 if (mapping->nr_bitmaps >= mapping->extensions)
2113 return -EINVAL;
2114
2115 next_bitmap = mapping->nr_bitmaps;
2116 mapping->bitmaps[next_bitmap] = kzalloc(mapping->bitmap_size,
2117 GFP_ATOMIC);
2118 if (!mapping->bitmaps[next_bitmap])
2119 return -ENOMEM;
2120
2121 mapping->nr_bitmaps++;
2122
2123 return 0;
2124}
2125
2126void arm_iommu_release_mapping(struct dma_iommu_mapping *mapping)
2127{
2128 if (mapping)
2129 kref_put(&mapping->kref, release_iommu_mapping);
2130}
2131EXPORT_SYMBOL_GPL(arm_iommu_release_mapping);
2132
2133static int __arm_iommu_attach_device(struct device *dev,
2134 struct dma_iommu_mapping *mapping)
2135{
2136 int err;
2137
2138 err = iommu_attach_device(mapping->domain, dev);
2139 if (err)
2140 return err;
2141
2142 kref_get(&mapping->kref);
2143 to_dma_iommu_mapping(dev) = mapping;
2144
2145 pr_debug("Attached IOMMU controller to %s device.\n", dev_name(dev));
2146 return 0;
2147}
2148
2149/**
2150 * arm_iommu_attach_device
2151 * @dev: valid struct device pointer
2152 * @mapping: io address space mapping structure (returned from
2153 * arm_iommu_create_mapping)
2154 *
2155 * Attaches specified io address space mapping to the provided device.
2156 * This replaces the dma operations (dma_map_ops pointer) with the
2157 * IOMMU aware version.
2158 *
2159 * More than one client might be attached to the same io address space
2160 * mapping.
2161 */
2162int arm_iommu_attach_device(struct device *dev,
2163 struct dma_iommu_mapping *mapping)
2164{
2165 int err;
2166
2167 err = __arm_iommu_attach_device(dev, mapping);
2168 if (err)
2169 return err;
2170
2171 set_dma_ops(dev, &iommu_ops);
2172 return 0;
2173}
2174EXPORT_SYMBOL_GPL(arm_iommu_attach_device);
2175
2176/**
2177 * arm_iommu_detach_device
2178 * @dev: valid struct device pointer
2179 *
2180 * Detaches the provided device from a previously attached map.
2181 * This overwrites the dma_ops pointer with appropriate non-IOMMU ops.
2182 */
2183void arm_iommu_detach_device(struct device *dev)
2184{
2185 struct dma_iommu_mapping *mapping;
2186
2187 mapping = to_dma_iommu_mapping(dev);
2188 if (!mapping) {
2189 dev_warn(dev, "Not attached\n");
2190 return;
2191 }
2192
2193 iommu_detach_device(mapping->domain, dev);
2194 kref_put(&mapping->kref, release_iommu_mapping);
2195 to_dma_iommu_mapping(dev) = NULL;
2196 set_dma_ops(dev, arm_get_dma_map_ops(dev->archdata.dma_coherent));
2197
2198 pr_debug("Detached IOMMU controller from %s device.\n", dev_name(dev));
2199}
2200EXPORT_SYMBOL_GPL(arm_iommu_detach_device);
2201
2202static const struct dma_map_ops *arm_get_iommu_dma_map_ops(bool coherent)
2203{
2204 return coherent ? &iommu_coherent_ops : &iommu_ops;
2205}
2206
2207static bool arm_setup_iommu_dma_ops(struct device *dev, u64 dma_base, u64 size,
2208 const struct iommu_ops *iommu)
2209{
2210 struct dma_iommu_mapping *mapping;
2211
2212 if (!iommu)
2213 return false;
2214
2215 mapping = arm_iommu_create_mapping(dev->bus, dma_base, size);
2216 if (IS_ERR(mapping)) {
2217 pr_warn("Failed to create %llu-byte IOMMU mapping for device %s\n",
2218 size, dev_name(dev));
2219 return false;
2220 }
2221
2222 if (__arm_iommu_attach_device(dev, mapping)) {
2223 pr_warn("Failed to attached device %s to IOMMU_mapping\n",
2224 dev_name(dev));
2225 arm_iommu_release_mapping(mapping);
2226 return false;
2227 }
2228
2229 return true;
2230}
2231
2232static void arm_teardown_iommu_dma_ops(struct device *dev)
2233{
2234 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2235
2236 if (!mapping)
2237 return;
2238
2239 arm_iommu_detach_device(dev);
2240 arm_iommu_release_mapping(mapping);
2241}
2242
2243#else
2244
2245static bool arm_setup_iommu_dma_ops(struct device *dev, u64 dma_base, u64 size,
2246 const struct iommu_ops *iommu)
2247{
2248 return false;
2249}
2250
2251static void arm_teardown_iommu_dma_ops(struct device *dev) { }
2252
2253#define arm_get_iommu_dma_map_ops arm_get_dma_map_ops
2254
2255#endif /* CONFIG_ARM_DMA_USE_IOMMU */
2256
2257void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
2258 const struct iommu_ops *iommu, bool coherent)
2259{
2260 const struct dma_map_ops *dma_ops;
2261
2262 dev->archdata.dma_coherent = coherent;
2263#ifdef CONFIG_SWIOTLB
2264 dev->dma_coherent = coherent;
2265#endif
2266
2267 /*
2268 * Don't override the dma_ops if they have already been set. Ideally
2269 * this should be the only location where dma_ops are set, remove this
2270 * check when all other callers of set_dma_ops will have disappeared.
2271 */
2272 if (dev->dma_ops)
2273 return;
2274
2275 if (arm_setup_iommu_dma_ops(dev, dma_base, size, iommu))
2276 dma_ops = arm_get_iommu_dma_map_ops(coherent);
2277 else
2278 dma_ops = arm_get_dma_map_ops(coherent);
2279
2280 set_dma_ops(dev, dma_ops);
2281
2282#ifdef CONFIG_XEN
2283 if (xen_initial_domain())
2284 dev->dma_ops = &xen_swiotlb_dma_ops;
2285#endif
2286 dev->archdata.dma_ops_setup = true;
2287}
2288
2289void arch_teardown_dma_ops(struct device *dev)
2290{
2291 if (!dev->archdata.dma_ops_setup)
2292 return;
2293
2294 arm_teardown_iommu_dma_ops(dev);
2295 /* Let arch_setup_dma_ops() start again from scratch upon re-probe */
2296 set_dma_ops(dev, NULL);
2297}
2298
2299#ifdef CONFIG_SWIOTLB
2300void arch_sync_dma_for_device(phys_addr_t paddr, size_t size,
2301 enum dma_data_direction dir)
2302{
2303 __dma_page_cpu_to_dev(phys_to_page(paddr), paddr & (PAGE_SIZE - 1),
2304 size, dir);
2305}
2306
2307void arch_sync_dma_for_cpu(phys_addr_t paddr, size_t size,
2308 enum dma_data_direction dir)
2309{
2310 __dma_page_dev_to_cpu(phys_to_page(paddr), paddr & (PAGE_SIZE - 1),
2311 size, dir);
2312}
2313
2314void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
2315 gfp_t gfp, unsigned long attrs)
2316{
2317 return __dma_alloc(dev, size, dma_handle, gfp,
2318 __get_dma_pgprot(attrs, PAGE_KERNEL), false,
2319 attrs, __builtin_return_address(0));
2320}
2321
2322void arch_dma_free(struct device *dev, size_t size, void *cpu_addr,
2323 dma_addr_t dma_handle, unsigned long attrs)
2324{
2325 __arm_dma_free(dev, size, cpu_addr, dma_handle, attrs, false);
2326}
2327#endif /* CONFIG_SWIOTLB */