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