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/*
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/bootmem.h>
13#include <linux/module.h>
14#include <linux/mm.h>
15#include <linux/genalloc.h>
16#include <linux/gfp.h>
17#include <linux/errno.h>
18#include <linux/list.h>
19#include <linux/init.h>
20#include <linux/device.h>
21#include <linux/dma-mapping.h>
22#include <linux/dma-contiguous.h>
23#include <linux/highmem.h>
24#include <linux/memblock.h>
25#include <linux/slab.h>
26#include <linux/iommu.h>
27#include <linux/io.h>
28#include <linux/vmalloc.h>
29#include <linux/sizes.h>
30#include <linux/cma.h>
31
32#include <asm/memory.h>
33#include <asm/highmem.h>
34#include <asm/cacheflush.h>
35#include <asm/tlbflush.h>
36#include <asm/mach/arch.h>
37#include <asm/dma-iommu.h>
38#include <asm/mach/map.h>
39#include <asm/system_info.h>
40#include <asm/dma-contiguous.h>
41
42#include "dma.h"
43#include "mm.h"
44
45struct arm_dma_alloc_args {
46 struct device *dev;
47 size_t size;
48 gfp_t gfp;
49 pgprot_t prot;
50 const void *caller;
51 bool want_vaddr;
52 int coherent_flag;
53};
54
55struct arm_dma_free_args {
56 struct device *dev;
57 size_t size;
58 void *cpu_addr;
59 struct page *page;
60 bool want_vaddr;
61};
62
63#define NORMAL 0
64#define COHERENT 1
65
66struct arm_dma_allocator {
67 void *(*alloc)(struct arm_dma_alloc_args *args,
68 struct page **ret_page);
69 void (*free)(struct arm_dma_free_args *args);
70};
71
72struct arm_dma_buffer {
73 struct list_head list;
74 void *virt;
75 struct arm_dma_allocator *allocator;
76};
77
78static LIST_HEAD(arm_dma_bufs);
79static DEFINE_SPINLOCK(arm_dma_bufs_lock);
80
81static struct arm_dma_buffer *arm_dma_buffer_find(void *virt)
82{
83 struct arm_dma_buffer *buf, *found = NULL;
84 unsigned long flags;
85
86 spin_lock_irqsave(&arm_dma_bufs_lock, flags);
87 list_for_each_entry(buf, &arm_dma_bufs, list) {
88 if (buf->virt == virt) {
89 list_del(&buf->list);
90 found = buf;
91 break;
92 }
93 }
94 spin_unlock_irqrestore(&arm_dma_bufs_lock, flags);
95 return found;
96}
97
98/*
99 * The DMA API is built upon the notion of "buffer ownership". A buffer
100 * is either exclusively owned by the CPU (and therefore may be accessed
101 * by it) or exclusively owned by the DMA device. These helper functions
102 * represent the transitions between these two ownership states.
103 *
104 * Note, however, that on later ARMs, this notion does not work due to
105 * speculative prefetches. We model our approach on the assumption that
106 * the CPU does do speculative prefetches, which means we clean caches
107 * before transfers and delay cache invalidation until transfer completion.
108 *
109 */
110static void __dma_page_cpu_to_dev(struct page *, unsigned long,
111 size_t, enum dma_data_direction);
112static void __dma_page_dev_to_cpu(struct page *, unsigned long,
113 size_t, enum dma_data_direction);
114
115/**
116 * arm_dma_map_page - map a portion of a page for streaming DMA
117 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
118 * @page: page that buffer resides in
119 * @offset: offset into page for start of buffer
120 * @size: size of buffer to map
121 * @dir: DMA transfer direction
122 *
123 * Ensure that any data held in the cache is appropriately discarded
124 * or written back.
125 *
126 * The device owns this memory once this call has completed. The CPU
127 * can regain ownership by calling dma_unmap_page().
128 */
129static dma_addr_t arm_dma_map_page(struct device *dev, struct page *page,
130 unsigned long offset, size_t size, enum dma_data_direction dir,
131 unsigned long attrs)
132{
133 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
134 __dma_page_cpu_to_dev(page, offset, size, dir);
135 return pfn_to_dma(dev, page_to_pfn(page)) + offset;
136}
137
138static dma_addr_t arm_coherent_dma_map_page(struct device *dev, struct page *page,
139 unsigned long offset, size_t size, enum dma_data_direction dir,
140 unsigned long attrs)
141{
142 return pfn_to_dma(dev, page_to_pfn(page)) + offset;
143}
144
145/**
146 * arm_dma_unmap_page - unmap a buffer previously mapped through dma_map_page()
147 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
148 * @handle: DMA address of buffer
149 * @size: size of buffer (same as passed to dma_map_page)
150 * @dir: DMA transfer direction (same as passed to dma_map_page)
151 *
152 * Unmap a page streaming mode DMA translation. The handle and size
153 * must match what was provided in the previous dma_map_page() call.
154 * All other usages are undefined.
155 *
156 * After this call, reads by the CPU to the buffer are guaranteed to see
157 * whatever the device wrote there.
158 */
159static void arm_dma_unmap_page(struct device *dev, dma_addr_t handle,
160 size_t size, enum dma_data_direction dir, unsigned long attrs)
161{
162 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
163 __dma_page_dev_to_cpu(pfn_to_page(dma_to_pfn(dev, handle)),
164 handle & ~PAGE_MASK, size, dir);
165}
166
167static void arm_dma_sync_single_for_cpu(struct device *dev,
168 dma_addr_t handle, size_t size, enum dma_data_direction dir)
169{
170 unsigned int offset = handle & (PAGE_SIZE - 1);
171 struct page *page = pfn_to_page(dma_to_pfn(dev, handle-offset));
172 __dma_page_dev_to_cpu(page, offset, size, dir);
173}
174
175static void arm_dma_sync_single_for_device(struct device *dev,
176 dma_addr_t handle, size_t size, enum dma_data_direction dir)
177{
178 unsigned int offset = handle & (PAGE_SIZE - 1);
179 struct page *page = pfn_to_page(dma_to_pfn(dev, handle-offset));
180 __dma_page_cpu_to_dev(page, offset, size, dir);
181}
182
183static int arm_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
184{
185 return dma_addr == ARM_MAPPING_ERROR;
186}
187
188const struct dma_map_ops arm_dma_ops = {
189 .alloc = arm_dma_alloc,
190 .free = arm_dma_free,
191 .mmap = arm_dma_mmap,
192 .get_sgtable = arm_dma_get_sgtable,
193 .map_page = arm_dma_map_page,
194 .unmap_page = arm_dma_unmap_page,
195 .map_sg = arm_dma_map_sg,
196 .unmap_sg = arm_dma_unmap_sg,
197 .sync_single_for_cpu = arm_dma_sync_single_for_cpu,
198 .sync_single_for_device = arm_dma_sync_single_for_device,
199 .sync_sg_for_cpu = arm_dma_sync_sg_for_cpu,
200 .sync_sg_for_device = arm_dma_sync_sg_for_device,
201 .mapping_error = arm_dma_mapping_error,
202 .dma_supported = arm_dma_supported,
203};
204EXPORT_SYMBOL(arm_dma_ops);
205
206static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
207 dma_addr_t *handle, gfp_t gfp, unsigned long attrs);
208static void arm_coherent_dma_free(struct device *dev, size_t size, void *cpu_addr,
209 dma_addr_t handle, unsigned long attrs);
210static int arm_coherent_dma_mmap(struct device *dev, struct vm_area_struct *vma,
211 void *cpu_addr, dma_addr_t dma_addr, size_t size,
212 unsigned long attrs);
213
214const struct dma_map_ops arm_coherent_dma_ops = {
215 .alloc = arm_coherent_dma_alloc,
216 .free = arm_coherent_dma_free,
217 .mmap = arm_coherent_dma_mmap,
218 .get_sgtable = arm_dma_get_sgtable,
219 .map_page = arm_coherent_dma_map_page,
220 .map_sg = arm_dma_map_sg,
221 .mapping_error = arm_dma_mapping_error,
222 .dma_supported = arm_dma_supported,
223};
224EXPORT_SYMBOL(arm_coherent_dma_ops);
225
226static int __dma_supported(struct device *dev, u64 mask, bool warn)
227{
228 unsigned long max_dma_pfn;
229
230 /*
231 * If the mask allows for more memory than we can address,
232 * and we actually have that much memory, then we must
233 * indicate that DMA to this device is not supported.
234 */
235 if (sizeof(mask) != sizeof(dma_addr_t) &&
236 mask > (dma_addr_t)~0 &&
237 dma_to_pfn(dev, ~0) < max_pfn - 1) {
238 if (warn) {
239 dev_warn(dev, "Coherent DMA mask %#llx is larger than dma_addr_t allows\n",
240 mask);
241 dev_warn(dev, "Driver did not use or check the return value from dma_set_coherent_mask()?\n");
242 }
243 return 0;
244 }
245
246 max_dma_pfn = min(max_pfn, arm_dma_pfn_limit);
247
248 /*
249 * Translate the device's DMA mask to a PFN limit. This
250 * PFN number includes the page which we can DMA to.
251 */
252 if (dma_to_pfn(dev, mask) < max_dma_pfn) {
253 if (warn)
254 dev_warn(dev, "Coherent DMA mask %#llx (pfn %#lx-%#lx) covers a smaller range of system memory than the DMA zone pfn 0x0-%#lx\n",
255 mask,
256 dma_to_pfn(dev, 0), dma_to_pfn(dev, mask) + 1,
257 max_dma_pfn + 1);
258 return 0;
259 }
260
261 return 1;
262}
263
264static u64 get_coherent_dma_mask(struct device *dev)
265{
266 u64 mask = (u64)DMA_BIT_MASK(32);
267
268 if (dev) {
269 mask = dev->coherent_dma_mask;
270
271 /*
272 * Sanity check the DMA mask - it must be non-zero, and
273 * must be able to be satisfied by a DMA allocation.
274 */
275 if (mask == 0) {
276 dev_warn(dev, "coherent DMA mask is unset\n");
277 return 0;
278 }
279
280 if (!__dma_supported(dev, mask, true))
281 return 0;
282 }
283
284 return mask;
285}
286
287static void __dma_clear_buffer(struct page *page, size_t size, int coherent_flag)
288{
289 /*
290 * Ensure that the allocated pages are zeroed, and that any data
291 * lurking in the kernel direct-mapped region is invalidated.
292 */
293 if (PageHighMem(page)) {
294 phys_addr_t base = __pfn_to_phys(page_to_pfn(page));
295 phys_addr_t end = base + size;
296 while (size > 0) {
297 void *ptr = kmap_atomic(page);
298 memset(ptr, 0, PAGE_SIZE);
299 if (coherent_flag != COHERENT)
300 dmac_flush_range(ptr, ptr + PAGE_SIZE);
301 kunmap_atomic(ptr);
302 page++;
303 size -= PAGE_SIZE;
304 }
305 if (coherent_flag != COHERENT)
306 outer_flush_range(base, end);
307 } else {
308 void *ptr = page_address(page);
309 memset(ptr, 0, size);
310 if (coherent_flag != COHERENT) {
311 dmac_flush_range(ptr, ptr + size);
312 outer_flush_range(__pa(ptr), __pa(ptr) + size);
313 }
314 }
315}
316
317/*
318 * Allocate a DMA buffer for 'dev' of size 'size' using the
319 * specified gfp mask. Note that 'size' must be page aligned.
320 */
321static struct page *__dma_alloc_buffer(struct device *dev, size_t size,
322 gfp_t gfp, int coherent_flag)
323{
324 unsigned long order = get_order(size);
325 struct page *page, *p, *e;
326
327 page = alloc_pages(gfp, order);
328 if (!page)
329 return NULL;
330
331 /*
332 * Now split the huge page and free the excess pages
333 */
334 split_page(page, order);
335 for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
336 __free_page(p);
337
338 __dma_clear_buffer(page, size, coherent_flag);
339
340 return page;
341}
342
343/*
344 * Free a DMA buffer. 'size' must be page aligned.
345 */
346static void __dma_free_buffer(struct page *page, size_t size)
347{
348 struct page *e = page + (size >> PAGE_SHIFT);
349
350 while (page < e) {
351 __free_page(page);
352 page++;
353 }
354}
355
356static void *__alloc_from_contiguous(struct device *dev, size_t size,
357 pgprot_t prot, struct page **ret_page,
358 const void *caller, bool want_vaddr,
359 int coherent_flag, gfp_t gfp);
360
361static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
362 pgprot_t prot, struct page **ret_page,
363 const void *caller, bool want_vaddr);
364
365static void *
366__dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot,
367 const void *caller)
368{
369 /*
370 * DMA allocation can be mapped to user space, so lets
371 * set VM_USERMAP flags too.
372 */
373 return dma_common_contiguous_remap(page, size,
374 VM_ARM_DMA_CONSISTENT | VM_USERMAP,
375 prot, caller);
376}
377
378static void __dma_free_remap(void *cpu_addr, size_t size)
379{
380 dma_common_free_remap(cpu_addr, size,
381 VM_ARM_DMA_CONSISTENT | VM_USERMAP);
382}
383
384#define DEFAULT_DMA_COHERENT_POOL_SIZE SZ_256K
385static struct gen_pool *atomic_pool __ro_after_init;
386
387static size_t atomic_pool_size __initdata = DEFAULT_DMA_COHERENT_POOL_SIZE;
388
389static int __init early_coherent_pool(char *p)
390{
391 atomic_pool_size = memparse(p, &p);
392 return 0;
393}
394early_param("coherent_pool", early_coherent_pool);
395
396/*
397 * Initialise the coherent pool for atomic allocations.
398 */
399static int __init atomic_pool_init(void)
400{
401 pgprot_t prot = pgprot_dmacoherent(PAGE_KERNEL);
402 gfp_t gfp = GFP_KERNEL | GFP_DMA;
403 struct page *page;
404 void *ptr;
405
406 atomic_pool = gen_pool_create(PAGE_SHIFT, -1);
407 if (!atomic_pool)
408 goto out;
409 /*
410 * The atomic pool is only used for non-coherent allocations
411 * so we must pass NORMAL for coherent_flag.
412 */
413 if (dev_get_cma_area(NULL))
414 ptr = __alloc_from_contiguous(NULL, atomic_pool_size, prot,
415 &page, atomic_pool_init, true, NORMAL,
416 GFP_KERNEL);
417 else
418 ptr = __alloc_remap_buffer(NULL, atomic_pool_size, gfp, prot,
419 &page, atomic_pool_init, true);
420 if (ptr) {
421 int ret;
422
423 ret = gen_pool_add_virt(atomic_pool, (unsigned long)ptr,
424 page_to_phys(page),
425 atomic_pool_size, -1);
426 if (ret)
427 goto destroy_genpool;
428
429 gen_pool_set_algo(atomic_pool,
430 gen_pool_first_fit_order_align,
431 NULL);
432 pr_info("DMA: preallocated %zu KiB pool for atomic coherent allocations\n",
433 atomic_pool_size / 1024);
434 return 0;
435 }
436
437destroy_genpool:
438 gen_pool_destroy(atomic_pool);
439 atomic_pool = NULL;
440out:
441 pr_err("DMA: failed to allocate %zu KiB pool for atomic coherent allocation\n",
442 atomic_pool_size / 1024);
443 return -ENOMEM;
444}
445/*
446 * CMA is activated by core_initcall, so we must be called after it.
447 */
448postcore_initcall(atomic_pool_init);
449
450struct dma_contig_early_reserve {
451 phys_addr_t base;
452 unsigned long size;
453};
454
455static struct dma_contig_early_reserve dma_mmu_remap[MAX_CMA_AREAS] __initdata;
456
457static int dma_mmu_remap_num __initdata;
458
459void __init dma_contiguous_early_fixup(phys_addr_t base, unsigned long size)
460{
461 dma_mmu_remap[dma_mmu_remap_num].base = base;
462 dma_mmu_remap[dma_mmu_remap_num].size = size;
463 dma_mmu_remap_num++;
464}
465
466void __init dma_contiguous_remap(void)
467{
468 int i;
469 for (i = 0; i < dma_mmu_remap_num; i++) {
470 phys_addr_t start = dma_mmu_remap[i].base;
471 phys_addr_t end = start + dma_mmu_remap[i].size;
472 struct map_desc map;
473 unsigned long addr;
474
475 if (end > arm_lowmem_limit)
476 end = arm_lowmem_limit;
477 if (start >= end)
478 continue;
479
480 map.pfn = __phys_to_pfn(start);
481 map.virtual = __phys_to_virt(start);
482 map.length = end - start;
483 map.type = MT_MEMORY_DMA_READY;
484
485 /*
486 * Clear previous low-memory mapping to ensure that the
487 * TLB does not see any conflicting entries, then flush
488 * the TLB of the old entries before creating new mappings.
489 *
490 * This ensures that any speculatively loaded TLB entries
491 * (even though they may be rare) can not cause any problems,
492 * and ensures that this code is architecturally compliant.
493 */
494 for (addr = __phys_to_virt(start); addr < __phys_to_virt(end);
495 addr += PMD_SIZE)
496 pmd_clear(pmd_off_k(addr));
497
498 flush_tlb_kernel_range(__phys_to_virt(start),
499 __phys_to_virt(end));
500
501 iotable_init(&map, 1);
502 }
503}
504
505static int __dma_update_pte(pte_t *pte, pgtable_t token, unsigned long addr,
506 void *data)
507{
508 struct page *page = virt_to_page(addr);
509 pgprot_t prot = *(pgprot_t *)data;
510
511 set_pte_ext(pte, mk_pte(page, prot), 0);
512 return 0;
513}
514
515static void __dma_remap(struct page *page, size_t size, pgprot_t prot)
516{
517 unsigned long start = (unsigned long) page_address(page);
518 unsigned end = start + size;
519
520 apply_to_page_range(&init_mm, start, size, __dma_update_pte, &prot);
521 flush_tlb_kernel_range(start, end);
522}
523
524static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
525 pgprot_t prot, struct page **ret_page,
526 const void *caller, bool want_vaddr)
527{
528 struct page *page;
529 void *ptr = NULL;
530 /*
531 * __alloc_remap_buffer is only called when the device is
532 * non-coherent
533 */
534 page = __dma_alloc_buffer(dev, size, gfp, NORMAL);
535 if (!page)
536 return NULL;
537 if (!want_vaddr)
538 goto out;
539
540 ptr = __dma_alloc_remap(page, size, gfp, prot, caller);
541 if (!ptr) {
542 __dma_free_buffer(page, size);
543 return NULL;
544 }
545
546 out:
547 *ret_page = page;
548 return ptr;
549}
550
551static void *__alloc_from_pool(size_t size, struct page **ret_page)
552{
553 unsigned long val;
554 void *ptr = NULL;
555
556 if (!atomic_pool) {
557 WARN(1, "coherent pool not initialised!\n");
558 return NULL;
559 }
560
561 val = gen_pool_alloc(atomic_pool, size);
562 if (val) {
563 phys_addr_t phys = gen_pool_virt_to_phys(atomic_pool, val);
564
565 *ret_page = phys_to_page(phys);
566 ptr = (void *)val;
567 }
568
569 return ptr;
570}
571
572static bool __in_atomic_pool(void *start, size_t size)
573{
574 return addr_in_gen_pool(atomic_pool, (unsigned long)start, size);
575}
576
577static int __free_from_pool(void *start, size_t size)
578{
579 if (!__in_atomic_pool(start, size))
580 return 0;
581
582 gen_pool_free(atomic_pool, (unsigned long)start, size);
583
584 return 1;
585}
586
587static void *__alloc_from_contiguous(struct device *dev, size_t size,
588 pgprot_t prot, struct page **ret_page,
589 const void *caller, bool want_vaddr,
590 int coherent_flag, gfp_t gfp)
591{
592 unsigned long order = get_order(size);
593 size_t count = size >> PAGE_SHIFT;
594 struct page *page;
595 void *ptr = NULL;
596
597 page = dma_alloc_from_contiguous(dev, count, order, gfp);
598 if (!page)
599 return NULL;
600
601 __dma_clear_buffer(page, size, coherent_flag);
602
603 if (!want_vaddr)
604 goto out;
605
606 if (PageHighMem(page)) {
607 ptr = __dma_alloc_remap(page, size, GFP_KERNEL, prot, caller);
608 if (!ptr) {
609 dma_release_from_contiguous(dev, page, count);
610 return NULL;
611 }
612 } else {
613 __dma_remap(page, size, prot);
614 ptr = page_address(page);
615 }
616
617 out:
618 *ret_page = page;
619 return ptr;
620}
621
622static void __free_from_contiguous(struct device *dev, struct page *page,
623 void *cpu_addr, size_t size, bool want_vaddr)
624{
625 if (want_vaddr) {
626 if (PageHighMem(page))
627 __dma_free_remap(cpu_addr, size);
628 else
629 __dma_remap(page, size, PAGE_KERNEL);
630 }
631 dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT);
632}
633
634static inline pgprot_t __get_dma_pgprot(unsigned long attrs, pgprot_t prot)
635{
636 prot = (attrs & DMA_ATTR_WRITE_COMBINE) ?
637 pgprot_writecombine(prot) :
638 pgprot_dmacoherent(prot);
639 return prot;
640}
641
642static void *__alloc_simple_buffer(struct device *dev, size_t size, gfp_t gfp,
643 struct page **ret_page)
644{
645 struct page *page;
646 /* __alloc_simple_buffer is only called when the device is coherent */
647 page = __dma_alloc_buffer(dev, size, gfp, COHERENT);
648 if (!page)
649 return NULL;
650
651 *ret_page = page;
652 return page_address(page);
653}
654
655static void *simple_allocator_alloc(struct arm_dma_alloc_args *args,
656 struct page **ret_page)
657{
658 return __alloc_simple_buffer(args->dev, args->size, args->gfp,
659 ret_page);
660}
661
662static void simple_allocator_free(struct arm_dma_free_args *args)
663{
664 __dma_free_buffer(args->page, args->size);
665}
666
667static struct arm_dma_allocator simple_allocator = {
668 .alloc = simple_allocator_alloc,
669 .free = simple_allocator_free,
670};
671
672static void *cma_allocator_alloc(struct arm_dma_alloc_args *args,
673 struct page **ret_page)
674{
675 return __alloc_from_contiguous(args->dev, args->size, args->prot,
676 ret_page, args->caller,
677 args->want_vaddr, args->coherent_flag,
678 args->gfp);
679}
680
681static void cma_allocator_free(struct arm_dma_free_args *args)
682{
683 __free_from_contiguous(args->dev, args->page, args->cpu_addr,
684 args->size, args->want_vaddr);
685}
686
687static struct arm_dma_allocator cma_allocator = {
688 .alloc = cma_allocator_alloc,
689 .free = cma_allocator_free,
690};
691
692static void *pool_allocator_alloc(struct arm_dma_alloc_args *args,
693 struct page **ret_page)
694{
695 return __alloc_from_pool(args->size, ret_page);
696}
697
698static void pool_allocator_free(struct arm_dma_free_args *args)
699{
700 __free_from_pool(args->cpu_addr, args->size);
701}
702
703static struct arm_dma_allocator pool_allocator = {
704 .alloc = pool_allocator_alloc,
705 .free = pool_allocator_free,
706};
707
708static void *remap_allocator_alloc(struct arm_dma_alloc_args *args,
709 struct page **ret_page)
710{
711 return __alloc_remap_buffer(args->dev, args->size, args->gfp,
712 args->prot, ret_page, args->caller,
713 args->want_vaddr);
714}
715
716static void remap_allocator_free(struct arm_dma_free_args *args)
717{
718 if (args->want_vaddr)
719 __dma_free_remap(args->cpu_addr, args->size);
720
721 __dma_free_buffer(args->page, args->size);
722}
723
724static struct arm_dma_allocator remap_allocator = {
725 .alloc = remap_allocator_alloc,
726 .free = remap_allocator_free,
727};
728
729static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
730 gfp_t gfp, pgprot_t prot, bool is_coherent,
731 unsigned long attrs, const void *caller)
732{
733 u64 mask = get_coherent_dma_mask(dev);
734 struct page *page = NULL;
735 void *addr;
736 bool allowblock, cma;
737 struct arm_dma_buffer *buf;
738 struct arm_dma_alloc_args args = {
739 .dev = dev,
740 .size = PAGE_ALIGN(size),
741 .gfp = gfp,
742 .prot = prot,
743 .caller = caller,
744 .want_vaddr = ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) == 0),
745 .coherent_flag = is_coherent ? COHERENT : NORMAL,
746 };
747
748#ifdef CONFIG_DMA_API_DEBUG
749 u64 limit = (mask + 1) & ~mask;
750 if (limit && size >= limit) {
751 dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n",
752 size, mask);
753 return NULL;
754 }
755#endif
756
757 if (!mask)
758 return NULL;
759
760 buf = kzalloc(sizeof(*buf),
761 gfp & ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM));
762 if (!buf)
763 return NULL;
764
765 if (mask < 0xffffffffULL)
766 gfp |= GFP_DMA;
767
768 /*
769 * Following is a work-around (a.k.a. hack) to prevent pages
770 * with __GFP_COMP being passed to split_page() which cannot
771 * handle them. The real problem is that this flag probably
772 * should be 0 on ARM as it is not supported on this
773 * platform; see CONFIG_HUGETLBFS.
774 */
775 gfp &= ~(__GFP_COMP);
776 args.gfp = gfp;
777
778 *handle = ARM_MAPPING_ERROR;
779 allowblock = gfpflags_allow_blocking(gfp);
780 cma = allowblock ? dev_get_cma_area(dev) : false;
781
782 if (cma)
783 buf->allocator = &cma_allocator;
784 else if (is_coherent)
785 buf->allocator = &simple_allocator;
786 else if (allowblock)
787 buf->allocator = &remap_allocator;
788 else
789 buf->allocator = &pool_allocator;
790
791 addr = buf->allocator->alloc(&args, &page);
792
793 if (page) {
794 unsigned long flags;
795
796 *handle = pfn_to_dma(dev, page_to_pfn(page));
797 buf->virt = args.want_vaddr ? addr : page;
798
799 spin_lock_irqsave(&arm_dma_bufs_lock, flags);
800 list_add(&buf->list, &arm_dma_bufs);
801 spin_unlock_irqrestore(&arm_dma_bufs_lock, flags);
802 } else {
803 kfree(buf);
804 }
805
806 return args.want_vaddr ? addr : page;
807}
808
809/*
810 * Allocate DMA-coherent memory space and return both the kernel remapped
811 * virtual and bus address for that space.
812 */
813void *arm_dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
814 gfp_t gfp, unsigned long attrs)
815{
816 pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
817
818 return __dma_alloc(dev, size, handle, gfp, prot, false,
819 attrs, __builtin_return_address(0));
820}
821
822static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
823 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
824{
825 return __dma_alloc(dev, size, handle, gfp, PAGE_KERNEL, true,
826 attrs, __builtin_return_address(0));
827}
828
829static int __arm_dma_mmap(struct device *dev, struct vm_area_struct *vma,
830 void *cpu_addr, dma_addr_t dma_addr, size_t size,
831 unsigned long attrs)
832{
833 int ret;
834 unsigned long nr_vma_pages = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
835 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
836 unsigned long pfn = dma_to_pfn(dev, dma_addr);
837 unsigned long off = vma->vm_pgoff;
838
839 if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
840 return ret;
841
842 if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) {
843 ret = remap_pfn_range(vma, vma->vm_start,
844 pfn + off,
845 vma->vm_end - vma->vm_start,
846 vma->vm_page_prot);
847 }
848
849 return ret;
850}
851
852/*
853 * Create userspace mapping for the DMA-coherent memory.
854 */
855static int arm_coherent_dma_mmap(struct device *dev, struct vm_area_struct *vma,
856 void *cpu_addr, dma_addr_t dma_addr, size_t size,
857 unsigned long attrs)
858{
859 return __arm_dma_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
860}
861
862int arm_dma_mmap(struct device *dev, struct vm_area_struct *vma,
863 void *cpu_addr, dma_addr_t dma_addr, size_t size,
864 unsigned long attrs)
865{
866 vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot);
867 return __arm_dma_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
868}
869
870/*
871 * Free a buffer as defined by the above mapping.
872 */
873static void __arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
874 dma_addr_t handle, unsigned long attrs,
875 bool is_coherent)
876{
877 struct page *page = pfn_to_page(dma_to_pfn(dev, handle));
878 struct arm_dma_buffer *buf;
879 struct arm_dma_free_args args = {
880 .dev = dev,
881 .size = PAGE_ALIGN(size),
882 .cpu_addr = cpu_addr,
883 .page = page,
884 .want_vaddr = ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) == 0),
885 };
886
887 buf = arm_dma_buffer_find(cpu_addr);
888 if (WARN(!buf, "Freeing invalid buffer %p\n", cpu_addr))
889 return;
890
891 buf->allocator->free(&args);
892 kfree(buf);
893}
894
895void arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
896 dma_addr_t handle, unsigned long attrs)
897{
898 __arm_dma_free(dev, size, cpu_addr, handle, attrs, false);
899}
900
901static void arm_coherent_dma_free(struct device *dev, size_t size, void *cpu_addr,
902 dma_addr_t handle, unsigned long attrs)
903{
904 __arm_dma_free(dev, size, cpu_addr, handle, attrs, true);
905}
906
907/*
908 * The whole dma_get_sgtable() idea is fundamentally unsafe - it seems
909 * that the intention is to allow exporting memory allocated via the
910 * coherent DMA APIs through the dma_buf API, which only accepts a
911 * scattertable. This presents a couple of problems:
912 * 1. Not all memory allocated via the coherent DMA APIs is backed by
913 * a struct page
914 * 2. Passing coherent DMA memory into the streaming APIs is not allowed
915 * as we will try to flush the memory through a different alias to that
916 * actually being used (and the flushes are redundant.)
917 */
918int arm_dma_get_sgtable(struct device *dev, struct sg_table *sgt,
919 void *cpu_addr, dma_addr_t handle, size_t size,
920 unsigned long attrs)
921{
922 unsigned long pfn = dma_to_pfn(dev, handle);
923 struct page *page;
924 int ret;
925
926 /* If the PFN is not valid, we do not have a struct page */
927 if (!pfn_valid(pfn))
928 return -ENXIO;
929
930 page = pfn_to_page(pfn);
931
932 ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
933 if (unlikely(ret))
934 return ret;
935
936 sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
937 return 0;
938}
939
940static void dma_cache_maint_page(struct page *page, unsigned long offset,
941 size_t size, enum dma_data_direction dir,
942 void (*op)(const void *, size_t, int))
943{
944 unsigned long pfn;
945 size_t left = size;
946
947 pfn = page_to_pfn(page) + offset / PAGE_SIZE;
948 offset %= PAGE_SIZE;
949
950 /*
951 * A single sg entry may refer to multiple physically contiguous
952 * pages. But we still need to process highmem pages individually.
953 * If highmem is not configured then the bulk of this loop gets
954 * optimized out.
955 */
956 do {
957 size_t len = left;
958 void *vaddr;
959
960 page = pfn_to_page(pfn);
961
962 if (PageHighMem(page)) {
963 if (len + offset > PAGE_SIZE)
964 len = PAGE_SIZE - offset;
965
966 if (cache_is_vipt_nonaliasing()) {
967 vaddr = kmap_atomic(page);
968 op(vaddr + offset, len, dir);
969 kunmap_atomic(vaddr);
970 } else {
971 vaddr = kmap_high_get(page);
972 if (vaddr) {
973 op(vaddr + offset, len, dir);
974 kunmap_high(page);
975 }
976 }
977 } else {
978 vaddr = page_address(page) + offset;
979 op(vaddr, len, dir);
980 }
981 offset = 0;
982 pfn++;
983 left -= len;
984 } while (left);
985}
986
987/*
988 * Make an area consistent for devices.
989 * Note: Drivers should NOT use this function directly, as it will break
990 * platforms with CONFIG_DMABOUNCE.
991 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
992 */
993static void __dma_page_cpu_to_dev(struct page *page, unsigned long off,
994 size_t size, enum dma_data_direction dir)
995{
996 phys_addr_t paddr;
997
998 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
999
1000 paddr = page_to_phys(page) + off;
1001 if (dir == DMA_FROM_DEVICE) {
1002 outer_inv_range(paddr, paddr + size);
1003 } else {
1004 outer_clean_range(paddr, paddr + size);
1005 }
1006 /* FIXME: non-speculating: flush on bidirectional mappings? */
1007}
1008
1009static void __dma_page_dev_to_cpu(struct page *page, unsigned long off,
1010 size_t size, enum dma_data_direction dir)
1011{
1012 phys_addr_t paddr = page_to_phys(page) + off;
1013
1014 /* FIXME: non-speculating: not required */
1015 /* in any case, don't bother invalidating if DMA to device */
1016 if (dir != DMA_TO_DEVICE) {
1017 outer_inv_range(paddr, paddr + size);
1018
1019 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
1020 }
1021
1022 /*
1023 * Mark the D-cache clean for these pages to avoid extra flushing.
1024 */
1025 if (dir != DMA_TO_DEVICE && size >= PAGE_SIZE) {
1026 unsigned long pfn;
1027 size_t left = size;
1028
1029 pfn = page_to_pfn(page) + off / PAGE_SIZE;
1030 off %= PAGE_SIZE;
1031 if (off) {
1032 pfn++;
1033 left -= PAGE_SIZE - off;
1034 }
1035 while (left >= PAGE_SIZE) {
1036 page = pfn_to_page(pfn++);
1037 set_bit(PG_dcache_clean, &page->flags);
1038 left -= PAGE_SIZE;
1039 }
1040 }
1041}
1042
1043/**
1044 * arm_dma_map_sg - map a set of SG buffers for streaming mode DMA
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
1048 * @dir: DMA transfer direction
1049 *
1050 * Map a set of buffers described by scatterlist in streaming mode for DMA.
1051 * This is the scatter-gather version of the dma_map_single interface.
1052 * Here the scatter gather list elements are each tagged with the
1053 * appropriate dma address and length. They are obtained via
1054 * sg_dma_{address,length}.
1055 *
1056 * Device ownership issues as mentioned for dma_map_single are the same
1057 * here.
1058 */
1059int arm_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
1060 enum dma_data_direction dir, unsigned long attrs)
1061{
1062 const struct dma_map_ops *ops = get_dma_ops(dev);
1063 struct scatterlist *s;
1064 int i, j;
1065
1066 for_each_sg(sg, s, nents, i) {
1067#ifdef CONFIG_NEED_SG_DMA_LENGTH
1068 s->dma_length = s->length;
1069#endif
1070 s->dma_address = ops->map_page(dev, sg_page(s), s->offset,
1071 s->length, dir, attrs);
1072 if (dma_mapping_error(dev, s->dma_address))
1073 goto bad_mapping;
1074 }
1075 return nents;
1076
1077 bad_mapping:
1078 for_each_sg(sg, s, i, j)
1079 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
1080 return 0;
1081}
1082
1083/**
1084 * arm_dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
1085 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
1086 * @sg: list of buffers
1087 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
1088 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1089 *
1090 * Unmap a set of streaming mode DMA translations. Again, CPU access
1091 * rules concerning calls here are the same as for dma_unmap_single().
1092 */
1093void arm_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
1094 enum dma_data_direction dir, unsigned long attrs)
1095{
1096 const struct dma_map_ops *ops = get_dma_ops(dev);
1097 struct scatterlist *s;
1098
1099 int i;
1100
1101 for_each_sg(sg, s, nents, i)
1102 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
1103}
1104
1105/**
1106 * arm_dma_sync_sg_for_cpu
1107 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
1108 * @sg: list of buffers
1109 * @nents: number of buffers to map (returned from dma_map_sg)
1110 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1111 */
1112void arm_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1113 int nents, enum dma_data_direction dir)
1114{
1115 const struct dma_map_ops *ops = get_dma_ops(dev);
1116 struct scatterlist *s;
1117 int i;
1118
1119 for_each_sg(sg, s, nents, i)
1120 ops->sync_single_for_cpu(dev, sg_dma_address(s), s->length,
1121 dir);
1122}
1123
1124/**
1125 * arm_dma_sync_sg_for_device
1126 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
1127 * @sg: list of buffers
1128 * @nents: number of buffers to map (returned from dma_map_sg)
1129 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1130 */
1131void arm_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1132 int nents, enum dma_data_direction dir)
1133{
1134 const struct dma_map_ops *ops = get_dma_ops(dev);
1135 struct scatterlist *s;
1136 int i;
1137
1138 for_each_sg(sg, s, nents, i)
1139 ops->sync_single_for_device(dev, sg_dma_address(s), s->length,
1140 dir);
1141}
1142
1143/*
1144 * Return whether the given device DMA address mask can be supported
1145 * properly. For example, if your device can only drive the low 24-bits
1146 * during bus mastering, then you would pass 0x00ffffff as the mask
1147 * to this function.
1148 */
1149int arm_dma_supported(struct device *dev, u64 mask)
1150{
1151 return __dma_supported(dev, mask, false);
1152}
1153
1154#define PREALLOC_DMA_DEBUG_ENTRIES 4096
1155
1156static int __init dma_debug_do_init(void)
1157{
1158 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
1159 return 0;
1160}
1161core_initcall(dma_debug_do_init);
1162
1163#ifdef CONFIG_ARM_DMA_USE_IOMMU
1164
1165static int __dma_info_to_prot(enum dma_data_direction dir, unsigned long attrs)
1166{
1167 int prot = 0;
1168
1169 if (attrs & DMA_ATTR_PRIVILEGED)
1170 prot |= IOMMU_PRIV;
1171
1172 switch (dir) {
1173 case DMA_BIDIRECTIONAL:
1174 return prot | IOMMU_READ | IOMMU_WRITE;
1175 case DMA_TO_DEVICE:
1176 return prot | IOMMU_READ;
1177 case DMA_FROM_DEVICE:
1178 return prot | IOMMU_WRITE;
1179 default:
1180 return prot;
1181 }
1182}
1183
1184/* IOMMU */
1185
1186static int extend_iommu_mapping(struct dma_iommu_mapping *mapping);
1187
1188static inline dma_addr_t __alloc_iova(struct dma_iommu_mapping *mapping,
1189 size_t size)
1190{
1191 unsigned int order = get_order(size);
1192 unsigned int align = 0;
1193 unsigned int count, start;
1194 size_t mapping_size = mapping->bits << PAGE_SHIFT;
1195 unsigned long flags;
1196 dma_addr_t iova;
1197 int i;
1198
1199 if (order > CONFIG_ARM_DMA_IOMMU_ALIGNMENT)
1200 order = CONFIG_ARM_DMA_IOMMU_ALIGNMENT;
1201
1202 count = PAGE_ALIGN(size) >> PAGE_SHIFT;
1203 align = (1 << order) - 1;
1204
1205 spin_lock_irqsave(&mapping->lock, flags);
1206 for (i = 0; i < mapping->nr_bitmaps; i++) {
1207 start = bitmap_find_next_zero_area(mapping->bitmaps[i],
1208 mapping->bits, 0, count, align);
1209
1210 if (start > mapping->bits)
1211 continue;
1212
1213 bitmap_set(mapping->bitmaps[i], start, count);
1214 break;
1215 }
1216
1217 /*
1218 * No unused range found. Try to extend the existing mapping
1219 * and perform a second attempt to reserve an IO virtual
1220 * address range of size bytes.
1221 */
1222 if (i == mapping->nr_bitmaps) {
1223 if (extend_iommu_mapping(mapping)) {
1224 spin_unlock_irqrestore(&mapping->lock, flags);
1225 return ARM_MAPPING_ERROR;
1226 }
1227
1228 start = bitmap_find_next_zero_area(mapping->bitmaps[i],
1229 mapping->bits, 0, count, align);
1230
1231 if (start > mapping->bits) {
1232 spin_unlock_irqrestore(&mapping->lock, flags);
1233 return ARM_MAPPING_ERROR;
1234 }
1235
1236 bitmap_set(mapping->bitmaps[i], start, count);
1237 }
1238 spin_unlock_irqrestore(&mapping->lock, flags);
1239
1240 iova = mapping->base + (mapping_size * i);
1241 iova += start << PAGE_SHIFT;
1242
1243 return iova;
1244}
1245
1246static inline void __free_iova(struct dma_iommu_mapping *mapping,
1247 dma_addr_t addr, size_t size)
1248{
1249 unsigned int start, count;
1250 size_t mapping_size = mapping->bits << PAGE_SHIFT;
1251 unsigned long flags;
1252 dma_addr_t bitmap_base;
1253 u32 bitmap_index;
1254
1255 if (!size)
1256 return;
1257
1258 bitmap_index = (u32) (addr - mapping->base) / (u32) mapping_size;
1259 BUG_ON(addr < mapping->base || bitmap_index > mapping->extensions);
1260
1261 bitmap_base = mapping->base + mapping_size * bitmap_index;
1262
1263 start = (addr - bitmap_base) >> PAGE_SHIFT;
1264
1265 if (addr + size > bitmap_base + mapping_size) {
1266 /*
1267 * The address range to be freed reaches into the iova
1268 * range of the next bitmap. This should not happen as
1269 * we don't allow this in __alloc_iova (at the
1270 * moment).
1271 */
1272 BUG();
1273 } else
1274 count = size >> PAGE_SHIFT;
1275
1276 spin_lock_irqsave(&mapping->lock, flags);
1277 bitmap_clear(mapping->bitmaps[bitmap_index], start, count);
1278 spin_unlock_irqrestore(&mapping->lock, flags);
1279}
1280
1281/* We'll try 2M, 1M, 64K, and finally 4K; array must end with 0! */
1282static const int iommu_order_array[] = { 9, 8, 4, 0 };
1283
1284static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
1285 gfp_t gfp, unsigned long attrs,
1286 int coherent_flag)
1287{
1288 struct page **pages;
1289 int count = size >> PAGE_SHIFT;
1290 int array_size = count * sizeof(struct page *);
1291 int i = 0;
1292 int order_idx = 0;
1293
1294 if (array_size <= PAGE_SIZE)
1295 pages = kzalloc(array_size, GFP_KERNEL);
1296 else
1297 pages = vzalloc(array_size);
1298 if (!pages)
1299 return NULL;
1300
1301 if (attrs & DMA_ATTR_FORCE_CONTIGUOUS)
1302 {
1303 unsigned long order = get_order(size);
1304 struct page *page;
1305
1306 page = dma_alloc_from_contiguous(dev, count, order, gfp);
1307 if (!page)
1308 goto error;
1309
1310 __dma_clear_buffer(page, size, coherent_flag);
1311
1312 for (i = 0; i < count; i++)
1313 pages[i] = page + i;
1314
1315 return pages;
1316 }
1317
1318 /* Go straight to 4K chunks if caller says it's OK. */
1319 if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)
1320 order_idx = ARRAY_SIZE(iommu_order_array) - 1;
1321
1322 /*
1323 * IOMMU can map any pages, so himem can also be used here
1324 */
1325 gfp |= __GFP_NOWARN | __GFP_HIGHMEM;
1326
1327 while (count) {
1328 int j, order;
1329
1330 order = iommu_order_array[order_idx];
1331
1332 /* Drop down when we get small */
1333 if (__fls(count) < order) {
1334 order_idx++;
1335 continue;
1336 }
1337
1338 if (order) {
1339 /* See if it's easy to allocate a high-order chunk */
1340 pages[i] = alloc_pages(gfp | __GFP_NORETRY, order);
1341
1342 /* Go down a notch at first sign of pressure */
1343 if (!pages[i]) {
1344 order_idx++;
1345 continue;
1346 }
1347 } else {
1348 pages[i] = alloc_pages(gfp, 0);
1349 if (!pages[i])
1350 goto error;
1351 }
1352
1353 if (order) {
1354 split_page(pages[i], order);
1355 j = 1 << order;
1356 while (--j)
1357 pages[i + j] = pages[i] + j;
1358 }
1359
1360 __dma_clear_buffer(pages[i], PAGE_SIZE << order, coherent_flag);
1361 i += 1 << order;
1362 count -= 1 << order;
1363 }
1364
1365 return pages;
1366error:
1367 while (i--)
1368 if (pages[i])
1369 __free_pages(pages[i], 0);
1370 kvfree(pages);
1371 return NULL;
1372}
1373
1374static int __iommu_free_buffer(struct device *dev, struct page **pages,
1375 size_t size, unsigned long attrs)
1376{
1377 int count = size >> PAGE_SHIFT;
1378 int i;
1379
1380 if (attrs & DMA_ATTR_FORCE_CONTIGUOUS) {
1381 dma_release_from_contiguous(dev, pages[0], count);
1382 } else {
1383 for (i = 0; i < count; i++)
1384 if (pages[i])
1385 __free_pages(pages[i], 0);
1386 }
1387
1388 kvfree(pages);
1389 return 0;
1390}
1391
1392/*
1393 * Create a CPU mapping for a specified pages
1394 */
1395static void *
1396__iommu_alloc_remap(struct page **pages, size_t size, gfp_t gfp, pgprot_t prot,
1397 const void *caller)
1398{
1399 return dma_common_pages_remap(pages, size,
1400 VM_ARM_DMA_CONSISTENT | VM_USERMAP, prot, caller);
1401}
1402
1403/*
1404 * Create a mapping in device IO address space for specified pages
1405 */
1406static dma_addr_t
1407__iommu_create_mapping(struct device *dev, struct page **pages, size_t size,
1408 unsigned long attrs)
1409{
1410 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1411 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
1412 dma_addr_t dma_addr, iova;
1413 int i;
1414
1415 dma_addr = __alloc_iova(mapping, size);
1416 if (dma_addr == ARM_MAPPING_ERROR)
1417 return dma_addr;
1418
1419 iova = dma_addr;
1420 for (i = 0; i < count; ) {
1421 int ret;
1422
1423 unsigned int next_pfn = page_to_pfn(pages[i]) + 1;
1424 phys_addr_t phys = page_to_phys(pages[i]);
1425 unsigned int len, j;
1426
1427 for (j = i + 1; j < count; j++, next_pfn++)
1428 if (page_to_pfn(pages[j]) != next_pfn)
1429 break;
1430
1431 len = (j - i) << PAGE_SHIFT;
1432 ret = iommu_map(mapping->domain, iova, phys, len,
1433 __dma_info_to_prot(DMA_BIDIRECTIONAL, attrs));
1434 if (ret < 0)
1435 goto fail;
1436 iova += len;
1437 i = j;
1438 }
1439 return dma_addr;
1440fail:
1441 iommu_unmap(mapping->domain, dma_addr, iova-dma_addr);
1442 __free_iova(mapping, dma_addr, size);
1443 return ARM_MAPPING_ERROR;
1444}
1445
1446static int __iommu_remove_mapping(struct device *dev, dma_addr_t iova, size_t size)
1447{
1448 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1449
1450 /*
1451 * add optional in-page offset from iova to size and align
1452 * result to page size
1453 */
1454 size = PAGE_ALIGN((iova & ~PAGE_MASK) + size);
1455 iova &= PAGE_MASK;
1456
1457 iommu_unmap(mapping->domain, iova, size);
1458 __free_iova(mapping, iova, size);
1459 return 0;
1460}
1461
1462static struct page **__atomic_get_pages(void *addr)
1463{
1464 struct page *page;
1465 phys_addr_t phys;
1466
1467 phys = gen_pool_virt_to_phys(atomic_pool, (unsigned long)addr);
1468 page = phys_to_page(phys);
1469
1470 return (struct page **)page;
1471}
1472
1473static struct page **__iommu_get_pages(void *cpu_addr, unsigned long attrs)
1474{
1475 struct vm_struct *area;
1476
1477 if (__in_atomic_pool(cpu_addr, PAGE_SIZE))
1478 return __atomic_get_pages(cpu_addr);
1479
1480 if (attrs & DMA_ATTR_NO_KERNEL_MAPPING)
1481 return cpu_addr;
1482
1483 area = find_vm_area(cpu_addr);
1484 if (area && (area->flags & VM_ARM_DMA_CONSISTENT))
1485 return area->pages;
1486 return NULL;
1487}
1488
1489static void *__iommu_alloc_simple(struct device *dev, size_t size, gfp_t gfp,
1490 dma_addr_t *handle, int coherent_flag,
1491 unsigned long attrs)
1492{
1493 struct page *page;
1494 void *addr;
1495
1496 if (coherent_flag == COHERENT)
1497 addr = __alloc_simple_buffer(dev, size, gfp, &page);
1498 else
1499 addr = __alloc_from_pool(size, &page);
1500 if (!addr)
1501 return NULL;
1502
1503 *handle = __iommu_create_mapping(dev, &page, size, attrs);
1504 if (*handle == ARM_MAPPING_ERROR)
1505 goto err_mapping;
1506
1507 return addr;
1508
1509err_mapping:
1510 __free_from_pool(addr, size);
1511 return NULL;
1512}
1513
1514static void __iommu_free_atomic(struct device *dev, void *cpu_addr,
1515 dma_addr_t handle, size_t size, int coherent_flag)
1516{
1517 __iommu_remove_mapping(dev, handle, size);
1518 if (coherent_flag == COHERENT)
1519 __dma_free_buffer(virt_to_page(cpu_addr), size);
1520 else
1521 __free_from_pool(cpu_addr, size);
1522}
1523
1524static void *__arm_iommu_alloc_attrs(struct device *dev, size_t size,
1525 dma_addr_t *handle, gfp_t gfp, unsigned long attrs,
1526 int coherent_flag)
1527{
1528 pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
1529 struct page **pages;
1530 void *addr = NULL;
1531
1532 *handle = ARM_MAPPING_ERROR;
1533 size = PAGE_ALIGN(size);
1534
1535 if (coherent_flag == COHERENT || !gfpflags_allow_blocking(gfp))
1536 return __iommu_alloc_simple(dev, size, gfp, handle,
1537 coherent_flag, attrs);
1538
1539 /*
1540 * Following is a work-around (a.k.a. hack) to prevent pages
1541 * with __GFP_COMP being passed to split_page() which cannot
1542 * handle them. The real problem is that this flag probably
1543 * should be 0 on ARM as it is not supported on this
1544 * platform; see CONFIG_HUGETLBFS.
1545 */
1546 gfp &= ~(__GFP_COMP);
1547
1548 pages = __iommu_alloc_buffer(dev, size, gfp, attrs, coherent_flag);
1549 if (!pages)
1550 return NULL;
1551
1552 *handle = __iommu_create_mapping(dev, pages, size, attrs);
1553 if (*handle == ARM_MAPPING_ERROR)
1554 goto err_buffer;
1555
1556 if (attrs & DMA_ATTR_NO_KERNEL_MAPPING)
1557 return pages;
1558
1559 addr = __iommu_alloc_remap(pages, size, gfp, prot,
1560 __builtin_return_address(0));
1561 if (!addr)
1562 goto err_mapping;
1563
1564 return addr;
1565
1566err_mapping:
1567 __iommu_remove_mapping(dev, *handle, size);
1568err_buffer:
1569 __iommu_free_buffer(dev, pages, size, attrs);
1570 return NULL;
1571}
1572
1573static void *arm_iommu_alloc_attrs(struct device *dev, size_t size,
1574 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1575{
1576 return __arm_iommu_alloc_attrs(dev, size, handle, gfp, attrs, NORMAL);
1577}
1578
1579static void *arm_coherent_iommu_alloc_attrs(struct device *dev, size_t size,
1580 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1581{
1582 return __arm_iommu_alloc_attrs(dev, size, handle, gfp, attrs, COHERENT);
1583}
1584
1585static int __arm_iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
1586 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1587 unsigned long attrs)
1588{
1589 unsigned long uaddr = vma->vm_start;
1590 unsigned long usize = vma->vm_end - vma->vm_start;
1591 struct page **pages = __iommu_get_pages(cpu_addr, attrs);
1592 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
1593 unsigned long off = vma->vm_pgoff;
1594
1595 if (!pages)
1596 return -ENXIO;
1597
1598 if (off >= nr_pages || (usize >> PAGE_SHIFT) > nr_pages - off)
1599 return -ENXIO;
1600
1601 pages += off;
1602
1603 do {
1604 int ret = vm_insert_page(vma, uaddr, *pages++);
1605 if (ret) {
1606 pr_err("Remapping memory failed: %d\n", ret);
1607 return ret;
1608 }
1609 uaddr += PAGE_SIZE;
1610 usize -= PAGE_SIZE;
1611 } while (usize > 0);
1612
1613 return 0;
1614}
1615static int arm_iommu_mmap_attrs(struct device *dev,
1616 struct vm_area_struct *vma, void *cpu_addr,
1617 dma_addr_t dma_addr, size_t size, unsigned long attrs)
1618{
1619 vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot);
1620
1621 return __arm_iommu_mmap_attrs(dev, vma, cpu_addr, dma_addr, size, attrs);
1622}
1623
1624static int arm_coherent_iommu_mmap_attrs(struct device *dev,
1625 struct vm_area_struct *vma, void *cpu_addr,
1626 dma_addr_t dma_addr, size_t size, unsigned long attrs)
1627{
1628 return __arm_iommu_mmap_attrs(dev, vma, cpu_addr, dma_addr, size, attrs);
1629}
1630
1631/*
1632 * free a page as defined by the above mapping.
1633 * Must not be called with IRQs disabled.
1634 */
1635void __arm_iommu_free_attrs(struct device *dev, size_t size, void *cpu_addr,
1636 dma_addr_t handle, unsigned long attrs, int coherent_flag)
1637{
1638 struct page **pages;
1639 size = PAGE_ALIGN(size);
1640
1641 if (coherent_flag == COHERENT || __in_atomic_pool(cpu_addr, size)) {
1642 __iommu_free_atomic(dev, cpu_addr, handle, size, coherent_flag);
1643 return;
1644 }
1645
1646 pages = __iommu_get_pages(cpu_addr, attrs);
1647 if (!pages) {
1648 WARN(1, "trying to free invalid coherent area: %p\n", cpu_addr);
1649 return;
1650 }
1651
1652 if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) == 0) {
1653 dma_common_free_remap(cpu_addr, size,
1654 VM_ARM_DMA_CONSISTENT | VM_USERMAP);
1655 }
1656
1657 __iommu_remove_mapping(dev, handle, size);
1658 __iommu_free_buffer(dev, pages, size, attrs);
1659}
1660
1661void arm_iommu_free_attrs(struct device *dev, size_t size,
1662 void *cpu_addr, dma_addr_t handle, unsigned long attrs)
1663{
1664 __arm_iommu_free_attrs(dev, size, cpu_addr, handle, attrs, NORMAL);
1665}
1666
1667void arm_coherent_iommu_free_attrs(struct device *dev, size_t size,
1668 void *cpu_addr, dma_addr_t handle, unsigned long attrs)
1669{
1670 __arm_iommu_free_attrs(dev, size, cpu_addr, handle, attrs, COHERENT);
1671}
1672
1673static int arm_iommu_get_sgtable(struct device *dev, struct sg_table *sgt,
1674 void *cpu_addr, dma_addr_t dma_addr,
1675 size_t size, unsigned long attrs)
1676{
1677 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
1678 struct page **pages = __iommu_get_pages(cpu_addr, attrs);
1679
1680 if (!pages)
1681 return -ENXIO;
1682
1683 return sg_alloc_table_from_pages(sgt, pages, count, 0, size,
1684 GFP_KERNEL);
1685}
1686
1687/*
1688 * Map a part of the scatter-gather list into contiguous io address space
1689 */
1690static int __map_sg_chunk(struct device *dev, struct scatterlist *sg,
1691 size_t size, dma_addr_t *handle,
1692 enum dma_data_direction dir, unsigned long attrs,
1693 bool is_coherent)
1694{
1695 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1696 dma_addr_t iova, iova_base;
1697 int ret = 0;
1698 unsigned int count;
1699 struct scatterlist *s;
1700 int prot;
1701
1702 size = PAGE_ALIGN(size);
1703 *handle = ARM_MAPPING_ERROR;
1704
1705 iova_base = iova = __alloc_iova(mapping, size);
1706 if (iova == ARM_MAPPING_ERROR)
1707 return -ENOMEM;
1708
1709 for (count = 0, s = sg; count < (size >> PAGE_SHIFT); s = sg_next(s)) {
1710 phys_addr_t phys = page_to_phys(sg_page(s));
1711 unsigned int len = PAGE_ALIGN(s->offset + s->length);
1712
1713 if (!is_coherent && (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1714 __dma_page_cpu_to_dev(sg_page(s), s->offset, s->length, dir);
1715
1716 prot = __dma_info_to_prot(dir, attrs);
1717
1718 ret = iommu_map(mapping->domain, iova, phys, len, prot);
1719 if (ret < 0)
1720 goto fail;
1721 count += len >> PAGE_SHIFT;
1722 iova += len;
1723 }
1724 *handle = iova_base;
1725
1726 return 0;
1727fail:
1728 iommu_unmap(mapping->domain, iova_base, count * PAGE_SIZE);
1729 __free_iova(mapping, iova_base, size);
1730 return ret;
1731}
1732
1733static int __iommu_map_sg(struct device *dev, struct scatterlist *sg, int nents,
1734 enum dma_data_direction dir, unsigned long attrs,
1735 bool is_coherent)
1736{
1737 struct scatterlist *s = sg, *dma = sg, *start = sg;
1738 int i, count = 0;
1739 unsigned int offset = s->offset;
1740 unsigned int size = s->offset + s->length;
1741 unsigned int max = dma_get_max_seg_size(dev);
1742
1743 for (i = 1; i < nents; i++) {
1744 s = sg_next(s);
1745
1746 s->dma_address = ARM_MAPPING_ERROR;
1747 s->dma_length = 0;
1748
1749 if (s->offset || (size & ~PAGE_MASK) || size + s->length > max) {
1750 if (__map_sg_chunk(dev, start, size, &dma->dma_address,
1751 dir, attrs, is_coherent) < 0)
1752 goto bad_mapping;
1753
1754 dma->dma_address += offset;
1755 dma->dma_length = size - offset;
1756
1757 size = offset = s->offset;
1758 start = s;
1759 dma = sg_next(dma);
1760 count += 1;
1761 }
1762 size += s->length;
1763 }
1764 if (__map_sg_chunk(dev, start, size, &dma->dma_address, dir, attrs,
1765 is_coherent) < 0)
1766 goto bad_mapping;
1767
1768 dma->dma_address += offset;
1769 dma->dma_length = size - offset;
1770
1771 return count+1;
1772
1773bad_mapping:
1774 for_each_sg(sg, s, count, i)
1775 __iommu_remove_mapping(dev, sg_dma_address(s), sg_dma_len(s));
1776 return 0;
1777}
1778
1779/**
1780 * arm_coherent_iommu_map_sg - map a set of SG buffers for streaming mode DMA
1781 * @dev: valid struct device pointer
1782 * @sg: list of buffers
1783 * @nents: number of buffers to map
1784 * @dir: DMA transfer direction
1785 *
1786 * Map a set of i/o coherent buffers described by scatterlist in streaming
1787 * mode for DMA. The scatter gather list elements are merged together (if
1788 * possible) and tagged with the appropriate dma address and length. They are
1789 * obtained via sg_dma_{address,length}.
1790 */
1791int arm_coherent_iommu_map_sg(struct device *dev, struct scatterlist *sg,
1792 int nents, enum dma_data_direction dir, unsigned long attrs)
1793{
1794 return __iommu_map_sg(dev, sg, nents, dir, attrs, true);
1795}
1796
1797/**
1798 * arm_iommu_map_sg - map a set of SG buffers for streaming mode DMA
1799 * @dev: valid struct device pointer
1800 * @sg: list of buffers
1801 * @nents: number of buffers to map
1802 * @dir: DMA transfer direction
1803 *
1804 * Map a set of buffers described by scatterlist in streaming mode for DMA.
1805 * The scatter gather list elements are merged together (if possible) and
1806 * tagged with the appropriate dma address and length. They are obtained via
1807 * sg_dma_{address,length}.
1808 */
1809int arm_iommu_map_sg(struct device *dev, struct scatterlist *sg,
1810 int nents, enum dma_data_direction dir, unsigned long attrs)
1811{
1812 return __iommu_map_sg(dev, sg, nents, dir, attrs, false);
1813}
1814
1815static void __iommu_unmap_sg(struct device *dev, struct scatterlist *sg,
1816 int nents, enum dma_data_direction dir,
1817 unsigned long attrs, bool is_coherent)
1818{
1819 struct scatterlist *s;
1820 int i;
1821
1822 for_each_sg(sg, s, nents, i) {
1823 if (sg_dma_len(s))
1824 __iommu_remove_mapping(dev, sg_dma_address(s),
1825 sg_dma_len(s));
1826 if (!is_coherent && (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1827 __dma_page_dev_to_cpu(sg_page(s), s->offset,
1828 s->length, dir);
1829 }
1830}
1831
1832/**
1833 * arm_coherent_iommu_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
1834 * @dev: valid struct device pointer
1835 * @sg: list of buffers
1836 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
1837 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1838 *
1839 * Unmap a set of streaming mode DMA translations. Again, CPU access
1840 * rules concerning calls here are the same as for dma_unmap_single().
1841 */
1842void arm_coherent_iommu_unmap_sg(struct device *dev, struct scatterlist *sg,
1843 int nents, enum dma_data_direction dir,
1844 unsigned long attrs)
1845{
1846 __iommu_unmap_sg(dev, sg, nents, dir, attrs, true);
1847}
1848
1849/**
1850 * arm_iommu_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
1851 * @dev: valid struct device pointer
1852 * @sg: list of buffers
1853 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
1854 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1855 *
1856 * Unmap a set of streaming mode DMA translations. Again, CPU access
1857 * rules concerning calls here are the same as for dma_unmap_single().
1858 */
1859void arm_iommu_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
1860 enum dma_data_direction dir,
1861 unsigned long attrs)
1862{
1863 __iommu_unmap_sg(dev, sg, nents, dir, attrs, false);
1864}
1865
1866/**
1867 * arm_iommu_sync_sg_for_cpu
1868 * @dev: valid struct device pointer
1869 * @sg: list of buffers
1870 * @nents: number of buffers to map (returned from dma_map_sg)
1871 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1872 */
1873void arm_iommu_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1874 int nents, enum dma_data_direction dir)
1875{
1876 struct scatterlist *s;
1877 int i;
1878
1879 for_each_sg(sg, s, nents, i)
1880 __dma_page_dev_to_cpu(sg_page(s), s->offset, s->length, dir);
1881
1882}
1883
1884/**
1885 * arm_iommu_sync_sg_for_device
1886 * @dev: valid struct device pointer
1887 * @sg: list of buffers
1888 * @nents: number of buffers to map (returned from dma_map_sg)
1889 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1890 */
1891void arm_iommu_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1892 int nents, enum dma_data_direction dir)
1893{
1894 struct scatterlist *s;
1895 int i;
1896
1897 for_each_sg(sg, s, nents, i)
1898 __dma_page_cpu_to_dev(sg_page(s), s->offset, s->length, dir);
1899}
1900
1901
1902/**
1903 * arm_coherent_iommu_map_page
1904 * @dev: valid struct device pointer
1905 * @page: page that buffer resides in
1906 * @offset: offset into page for start of buffer
1907 * @size: size of buffer to map
1908 * @dir: DMA transfer direction
1909 *
1910 * Coherent IOMMU aware version of arm_dma_map_page()
1911 */
1912static dma_addr_t arm_coherent_iommu_map_page(struct device *dev, struct page *page,
1913 unsigned long offset, size_t size, enum dma_data_direction dir,
1914 unsigned long attrs)
1915{
1916 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1917 dma_addr_t dma_addr;
1918 int ret, prot, len = PAGE_ALIGN(size + offset);
1919
1920 dma_addr = __alloc_iova(mapping, len);
1921 if (dma_addr == ARM_MAPPING_ERROR)
1922 return dma_addr;
1923
1924 prot = __dma_info_to_prot(dir, attrs);
1925
1926 ret = iommu_map(mapping->domain, dma_addr, page_to_phys(page), len, prot);
1927 if (ret < 0)
1928 goto fail;
1929
1930 return dma_addr + offset;
1931fail:
1932 __free_iova(mapping, dma_addr, len);
1933 return ARM_MAPPING_ERROR;
1934}
1935
1936/**
1937 * arm_iommu_map_page
1938 * @dev: valid struct device pointer
1939 * @page: page that buffer resides in
1940 * @offset: offset into page for start of buffer
1941 * @size: size of buffer to map
1942 * @dir: DMA transfer direction
1943 *
1944 * IOMMU aware version of arm_dma_map_page()
1945 */
1946static dma_addr_t arm_iommu_map_page(struct device *dev, struct page *page,
1947 unsigned long offset, size_t size, enum dma_data_direction dir,
1948 unsigned long attrs)
1949{
1950 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1951 __dma_page_cpu_to_dev(page, offset, size, dir);
1952
1953 return arm_coherent_iommu_map_page(dev, page, offset, size, dir, attrs);
1954}
1955
1956/**
1957 * arm_coherent_iommu_unmap_page
1958 * @dev: valid struct device pointer
1959 * @handle: DMA address of buffer
1960 * @size: size of buffer (same as passed to dma_map_page)
1961 * @dir: DMA transfer direction (same as passed to dma_map_page)
1962 *
1963 * Coherent IOMMU aware version of arm_dma_unmap_page()
1964 */
1965static void arm_coherent_iommu_unmap_page(struct device *dev, dma_addr_t handle,
1966 size_t size, enum dma_data_direction dir, unsigned long attrs)
1967{
1968 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1969 dma_addr_t iova = handle & PAGE_MASK;
1970 int offset = handle & ~PAGE_MASK;
1971 int len = PAGE_ALIGN(size + offset);
1972
1973 if (!iova)
1974 return;
1975
1976 iommu_unmap(mapping->domain, iova, len);
1977 __free_iova(mapping, iova, len);
1978}
1979
1980/**
1981 * arm_iommu_unmap_page
1982 * @dev: valid struct device pointer
1983 * @handle: DMA address of buffer
1984 * @size: size of buffer (same as passed to dma_map_page)
1985 * @dir: DMA transfer direction (same as passed to dma_map_page)
1986 *
1987 * IOMMU aware version of arm_dma_unmap_page()
1988 */
1989static void arm_iommu_unmap_page(struct device *dev, dma_addr_t handle,
1990 size_t size, enum dma_data_direction dir, unsigned long attrs)
1991{
1992 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1993 dma_addr_t iova = handle & PAGE_MASK;
1994 struct page *page = phys_to_page(iommu_iova_to_phys(mapping->domain, iova));
1995 int offset = handle & ~PAGE_MASK;
1996 int len = PAGE_ALIGN(size + offset);
1997
1998 if (!iova)
1999 return;
2000
2001 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
2002 __dma_page_dev_to_cpu(page, offset, size, dir);
2003
2004 iommu_unmap(mapping->domain, iova, len);
2005 __free_iova(mapping, iova, len);
2006}
2007
2008/**
2009 * arm_iommu_map_resource - map a device resource for DMA
2010 * @dev: valid struct device pointer
2011 * @phys_addr: physical address of resource
2012 * @size: size of resource to map
2013 * @dir: DMA transfer direction
2014 */
2015static dma_addr_t arm_iommu_map_resource(struct device *dev,
2016 phys_addr_t phys_addr, size_t size,
2017 enum dma_data_direction dir, unsigned long attrs)
2018{
2019 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2020 dma_addr_t dma_addr;
2021 int ret, prot;
2022 phys_addr_t addr = phys_addr & PAGE_MASK;
2023 unsigned int offset = phys_addr & ~PAGE_MASK;
2024 size_t len = PAGE_ALIGN(size + offset);
2025
2026 dma_addr = __alloc_iova(mapping, len);
2027 if (dma_addr == ARM_MAPPING_ERROR)
2028 return dma_addr;
2029
2030 prot = __dma_info_to_prot(dir, attrs) | IOMMU_MMIO;
2031
2032 ret = iommu_map(mapping->domain, dma_addr, addr, len, prot);
2033 if (ret < 0)
2034 goto fail;
2035
2036 return dma_addr + offset;
2037fail:
2038 __free_iova(mapping, dma_addr, len);
2039 return ARM_MAPPING_ERROR;
2040}
2041
2042/**
2043 * arm_iommu_unmap_resource - unmap a device DMA resource
2044 * @dev: valid struct device pointer
2045 * @dma_handle: DMA address to resource
2046 * @size: size of resource to map
2047 * @dir: DMA transfer direction
2048 */
2049static void arm_iommu_unmap_resource(struct device *dev, dma_addr_t dma_handle,
2050 size_t size, enum dma_data_direction dir,
2051 unsigned long attrs)
2052{
2053 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2054 dma_addr_t iova = dma_handle & PAGE_MASK;
2055 unsigned int offset = dma_handle & ~PAGE_MASK;
2056 size_t len = PAGE_ALIGN(size + offset);
2057
2058 if (!iova)
2059 return;
2060
2061 iommu_unmap(mapping->domain, iova, len);
2062 __free_iova(mapping, iova, len);
2063}
2064
2065static void arm_iommu_sync_single_for_cpu(struct device *dev,
2066 dma_addr_t handle, size_t size, enum dma_data_direction dir)
2067{
2068 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2069 dma_addr_t iova = handle & PAGE_MASK;
2070 struct page *page = phys_to_page(iommu_iova_to_phys(mapping->domain, iova));
2071 unsigned int offset = handle & ~PAGE_MASK;
2072
2073 if (!iova)
2074 return;
2075
2076 __dma_page_dev_to_cpu(page, offset, size, dir);
2077}
2078
2079static void arm_iommu_sync_single_for_device(struct device *dev,
2080 dma_addr_t handle, size_t size, enum dma_data_direction dir)
2081{
2082 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2083 dma_addr_t iova = handle & PAGE_MASK;
2084 struct page *page = phys_to_page(iommu_iova_to_phys(mapping->domain, iova));
2085 unsigned int offset = handle & ~PAGE_MASK;
2086
2087 if (!iova)
2088 return;
2089
2090 __dma_page_cpu_to_dev(page, offset, size, dir);
2091}
2092
2093const struct dma_map_ops iommu_ops = {
2094 .alloc = arm_iommu_alloc_attrs,
2095 .free = arm_iommu_free_attrs,
2096 .mmap = arm_iommu_mmap_attrs,
2097 .get_sgtable = arm_iommu_get_sgtable,
2098
2099 .map_page = arm_iommu_map_page,
2100 .unmap_page = arm_iommu_unmap_page,
2101 .sync_single_for_cpu = arm_iommu_sync_single_for_cpu,
2102 .sync_single_for_device = arm_iommu_sync_single_for_device,
2103
2104 .map_sg = arm_iommu_map_sg,
2105 .unmap_sg = arm_iommu_unmap_sg,
2106 .sync_sg_for_cpu = arm_iommu_sync_sg_for_cpu,
2107 .sync_sg_for_device = arm_iommu_sync_sg_for_device,
2108
2109 .map_resource = arm_iommu_map_resource,
2110 .unmap_resource = arm_iommu_unmap_resource,
2111
2112 .mapping_error = arm_dma_mapping_error,
2113 .dma_supported = arm_dma_supported,
2114};
2115
2116const struct dma_map_ops iommu_coherent_ops = {
2117 .alloc = arm_coherent_iommu_alloc_attrs,
2118 .free = arm_coherent_iommu_free_attrs,
2119 .mmap = arm_coherent_iommu_mmap_attrs,
2120 .get_sgtable = arm_iommu_get_sgtable,
2121
2122 .map_page = arm_coherent_iommu_map_page,
2123 .unmap_page = arm_coherent_iommu_unmap_page,
2124
2125 .map_sg = arm_coherent_iommu_map_sg,
2126 .unmap_sg = arm_coherent_iommu_unmap_sg,
2127
2128 .map_resource = arm_iommu_map_resource,
2129 .unmap_resource = arm_iommu_unmap_resource,
2130
2131 .mapping_error = arm_dma_mapping_error,
2132 .dma_supported = arm_dma_supported,
2133};
2134
2135/**
2136 * arm_iommu_create_mapping
2137 * @bus: pointer to the bus holding the client device (for IOMMU calls)
2138 * @base: start address of the valid IO address space
2139 * @size: maximum size of the valid IO address space
2140 *
2141 * Creates a mapping structure which holds information about used/unused
2142 * IO address ranges, which is required to perform memory allocation and
2143 * mapping with IOMMU aware functions.
2144 *
2145 * The client device need to be attached to the mapping with
2146 * arm_iommu_attach_device function.
2147 */
2148struct dma_iommu_mapping *
2149arm_iommu_create_mapping(struct bus_type *bus, dma_addr_t base, u64 size)
2150{
2151 unsigned int bits = size >> PAGE_SHIFT;
2152 unsigned int bitmap_size = BITS_TO_LONGS(bits) * sizeof(long);
2153 struct dma_iommu_mapping *mapping;
2154 int extensions = 1;
2155 int err = -ENOMEM;
2156
2157 /* currently only 32-bit DMA address space is supported */
2158 if (size > DMA_BIT_MASK(32) + 1)
2159 return ERR_PTR(-ERANGE);
2160
2161 if (!bitmap_size)
2162 return ERR_PTR(-EINVAL);
2163
2164 if (bitmap_size > PAGE_SIZE) {
2165 extensions = bitmap_size / PAGE_SIZE;
2166 bitmap_size = PAGE_SIZE;
2167 }
2168
2169 mapping = kzalloc(sizeof(struct dma_iommu_mapping), GFP_KERNEL);
2170 if (!mapping)
2171 goto err;
2172
2173 mapping->bitmap_size = bitmap_size;
2174 mapping->bitmaps = kzalloc(extensions * sizeof(unsigned long *),
2175 GFP_KERNEL);
2176 if (!mapping->bitmaps)
2177 goto err2;
2178
2179 mapping->bitmaps[0] = kzalloc(bitmap_size, GFP_KERNEL);
2180 if (!mapping->bitmaps[0])
2181 goto err3;
2182
2183 mapping->nr_bitmaps = 1;
2184 mapping->extensions = extensions;
2185 mapping->base = base;
2186 mapping->bits = BITS_PER_BYTE * bitmap_size;
2187
2188 spin_lock_init(&mapping->lock);
2189
2190 mapping->domain = iommu_domain_alloc(bus);
2191 if (!mapping->domain)
2192 goto err4;
2193
2194 kref_init(&mapping->kref);
2195 return mapping;
2196err4:
2197 kfree(mapping->bitmaps[0]);
2198err3:
2199 kfree(mapping->bitmaps);
2200err2:
2201 kfree(mapping);
2202err:
2203 return ERR_PTR(err);
2204}
2205EXPORT_SYMBOL_GPL(arm_iommu_create_mapping);
2206
2207static void release_iommu_mapping(struct kref *kref)
2208{
2209 int i;
2210 struct dma_iommu_mapping *mapping =
2211 container_of(kref, struct dma_iommu_mapping, kref);
2212
2213 iommu_domain_free(mapping->domain);
2214 for (i = 0; i < mapping->nr_bitmaps; i++)
2215 kfree(mapping->bitmaps[i]);
2216 kfree(mapping->bitmaps);
2217 kfree(mapping);
2218}
2219
2220static int extend_iommu_mapping(struct dma_iommu_mapping *mapping)
2221{
2222 int next_bitmap;
2223
2224 if (mapping->nr_bitmaps >= mapping->extensions)
2225 return -EINVAL;
2226
2227 next_bitmap = mapping->nr_bitmaps;
2228 mapping->bitmaps[next_bitmap] = kzalloc(mapping->bitmap_size,
2229 GFP_ATOMIC);
2230 if (!mapping->bitmaps[next_bitmap])
2231 return -ENOMEM;
2232
2233 mapping->nr_bitmaps++;
2234
2235 return 0;
2236}
2237
2238void arm_iommu_release_mapping(struct dma_iommu_mapping *mapping)
2239{
2240 if (mapping)
2241 kref_put(&mapping->kref, release_iommu_mapping);
2242}
2243EXPORT_SYMBOL_GPL(arm_iommu_release_mapping);
2244
2245static int __arm_iommu_attach_device(struct device *dev,
2246 struct dma_iommu_mapping *mapping)
2247{
2248 int err;
2249
2250 err = iommu_attach_device(mapping->domain, dev);
2251 if (err)
2252 return err;
2253
2254 kref_get(&mapping->kref);
2255 to_dma_iommu_mapping(dev) = mapping;
2256
2257 pr_debug("Attached IOMMU controller to %s device.\n", dev_name(dev));
2258 return 0;
2259}
2260
2261/**
2262 * arm_iommu_attach_device
2263 * @dev: valid struct device pointer
2264 * @mapping: io address space mapping structure (returned from
2265 * arm_iommu_create_mapping)
2266 *
2267 * Attaches specified io address space mapping to the provided device.
2268 * This replaces the dma operations (dma_map_ops pointer) with the
2269 * IOMMU aware version.
2270 *
2271 * More than one client might be attached to the same io address space
2272 * mapping.
2273 */
2274int arm_iommu_attach_device(struct device *dev,
2275 struct dma_iommu_mapping *mapping)
2276{
2277 int err;
2278
2279 err = __arm_iommu_attach_device(dev, mapping);
2280 if (err)
2281 return err;
2282
2283 set_dma_ops(dev, &iommu_ops);
2284 return 0;
2285}
2286EXPORT_SYMBOL_GPL(arm_iommu_attach_device);
2287
2288/**
2289 * arm_iommu_detach_device
2290 * @dev: valid struct device pointer
2291 *
2292 * Detaches the provided device from a previously attached map.
2293 * This voids the dma operations (dma_map_ops pointer)
2294 */
2295void arm_iommu_detach_device(struct device *dev)
2296{
2297 struct dma_iommu_mapping *mapping;
2298
2299 mapping = to_dma_iommu_mapping(dev);
2300 if (!mapping) {
2301 dev_warn(dev, "Not attached\n");
2302 return;
2303 }
2304
2305 iommu_detach_device(mapping->domain, dev);
2306 kref_put(&mapping->kref, release_iommu_mapping);
2307 to_dma_iommu_mapping(dev) = NULL;
2308 set_dma_ops(dev, NULL);
2309
2310 pr_debug("Detached IOMMU controller from %s device.\n", dev_name(dev));
2311}
2312EXPORT_SYMBOL_GPL(arm_iommu_detach_device);
2313
2314static const struct dma_map_ops *arm_get_iommu_dma_map_ops(bool coherent)
2315{
2316 return coherent ? &iommu_coherent_ops : &iommu_ops;
2317}
2318
2319static bool arm_setup_iommu_dma_ops(struct device *dev, u64 dma_base, u64 size,
2320 const struct iommu_ops *iommu)
2321{
2322 struct dma_iommu_mapping *mapping;
2323
2324 if (!iommu)
2325 return false;
2326
2327 mapping = arm_iommu_create_mapping(dev->bus, dma_base, size);
2328 if (IS_ERR(mapping)) {
2329 pr_warn("Failed to create %llu-byte IOMMU mapping for device %s\n",
2330 size, dev_name(dev));
2331 return false;
2332 }
2333
2334 if (__arm_iommu_attach_device(dev, mapping)) {
2335 pr_warn("Failed to attached device %s to IOMMU_mapping\n",
2336 dev_name(dev));
2337 arm_iommu_release_mapping(mapping);
2338 return false;
2339 }
2340
2341 return true;
2342}
2343
2344static void arm_teardown_iommu_dma_ops(struct device *dev)
2345{
2346 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2347
2348 if (!mapping)
2349 return;
2350
2351 arm_iommu_detach_device(dev);
2352 arm_iommu_release_mapping(mapping);
2353}
2354
2355#else
2356
2357static bool arm_setup_iommu_dma_ops(struct device *dev, u64 dma_base, u64 size,
2358 const struct iommu_ops *iommu)
2359{
2360 return false;
2361}
2362
2363static void arm_teardown_iommu_dma_ops(struct device *dev) { }
2364
2365#define arm_get_iommu_dma_map_ops arm_get_dma_map_ops
2366
2367#endif /* CONFIG_ARM_DMA_USE_IOMMU */
2368
2369static const struct dma_map_ops *arm_get_dma_map_ops(bool coherent)
2370{
2371 return coherent ? &arm_coherent_dma_ops : &arm_dma_ops;
2372}
2373
2374void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
2375 const struct iommu_ops *iommu, bool coherent)
2376{
2377 const struct dma_map_ops *dma_ops;
2378
2379 dev->archdata.dma_coherent = coherent;
2380
2381 /*
2382 * Don't override the dma_ops if they have already been set. Ideally
2383 * this should be the only location where dma_ops are set, remove this
2384 * check when all other callers of set_dma_ops will have disappeared.
2385 */
2386 if (dev->dma_ops)
2387 return;
2388
2389 if (arm_setup_iommu_dma_ops(dev, dma_base, size, iommu))
2390 dma_ops = arm_get_iommu_dma_map_ops(coherent);
2391 else
2392 dma_ops = arm_get_dma_map_ops(coherent);
2393
2394 set_dma_ops(dev, dma_ops);
2395
2396#ifdef CONFIG_XEN
2397 if (xen_initial_domain()) {
2398 dev->archdata.dev_dma_ops = dev->dma_ops;
2399 dev->dma_ops = xen_dma_ops;
2400 }
2401#endif
2402 dev->archdata.dma_ops_setup = true;
2403}
2404
2405void arch_teardown_dma_ops(struct device *dev)
2406{
2407 if (!dev->archdata.dma_ops_setup)
2408 return;
2409
2410 arm_teardown_iommu_dma_ops(dev);
2411}