2 * linux/arch/arm/mm/dma-mapping.c
4 * Copyright (C) 2000-2004 Russell King
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.
10 * DMA uncached mapping support.
12 #include <linux/module.h>
14 #include <linux/genalloc.h>
15 #include <linux/gfp.h>
16 #include <linux/errno.h>
17 #include <linux/list.h>
18 #include <linux/init.h>
19 #include <linux/device.h>
20 #include <linux/dma-mapping.h>
21 #include <linux/dma-contiguous.h>
22 #include <linux/highmem.h>
23 #include <linux/memblock.h>
24 #include <linux/slab.h>
25 #include <linux/iommu.h>
27 #include <linux/vmalloc.h>
28 #include <linux/sizes.h>
29 #include <linux/cma.h>
31 #include <asm/memory.h>
32 #include <asm/highmem.h>
33 #include <asm/cacheflush.h>
34 #include <asm/tlbflush.h>
35 #include <asm/mach/arch.h>
36 #include <asm/dma-iommu.h>
37 #include <asm/mach/map.h>
38 #include <asm/system_info.h>
39 #include <asm/dma-contiguous.h>
44 struct arm_dma_alloc_args {
54 struct arm_dma_free_args {
65 struct 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);
71 struct arm_dma_buffer {
72 struct list_head list;
74 struct arm_dma_allocator *allocator;
77 static LIST_HEAD(arm_dma_bufs);
78 static DEFINE_SPINLOCK(arm_dma_bufs_lock);
80 static struct arm_dma_buffer *arm_dma_buffer_find(void *virt)
82 struct arm_dma_buffer *buf, *found = NULL;
85 spin_lock_irqsave(&arm_dma_bufs_lock, flags);
86 list_for_each_entry(buf, &arm_dma_bufs, list) {
87 if (buf->virt == virt) {
93 spin_unlock_irqrestore(&arm_dma_bufs_lock, flags);
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.
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.
109 static void __dma_page_cpu_to_dev(struct page *, unsigned long,
110 size_t, enum dma_data_direction);
111 static void __dma_page_dev_to_cpu(struct page *, unsigned long,
112 size_t, enum dma_data_direction);
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
122 * Ensure that any data held in the cache is appropriately discarded
125 * The device owns this memory once this call has completed. The CPU
126 * can regain ownership by calling dma_unmap_page().
128 static 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,
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;
137 static 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,
141 return pfn_to_dma(dev, page_to_pfn(page)) + offset;
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)
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.
155 * After this call, reads by the CPU to the buffer are guaranteed to see
156 * whatever the device wrote there.
158 static void arm_dma_unmap_page(struct device *dev, dma_addr_t handle,
159 size_t size, enum dma_data_direction dir, unsigned long attrs)
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);
166 static void arm_dma_sync_single_for_cpu(struct device *dev,
167 dma_addr_t handle, size_t size, enum dma_data_direction dir)
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);
174 static void arm_dma_sync_single_for_device(struct device *dev,
175 dma_addr_t handle, size_t size, enum dma_data_direction dir)
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);
182 static int arm_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
184 return dma_addr == ARM_MAPPING_ERROR;
187 const struct dma_map_ops arm_dma_ops = {
188 .alloc = arm_dma_alloc,
189 .free = arm_dma_free,
190 .mmap = arm_dma_mmap,
191 .get_sgtable = arm_dma_get_sgtable,
192 .map_page = arm_dma_map_page,
193 .unmap_page = arm_dma_unmap_page,
194 .map_sg = arm_dma_map_sg,
195 .unmap_sg = arm_dma_unmap_sg,
196 .sync_single_for_cpu = arm_dma_sync_single_for_cpu,
197 .sync_single_for_device = arm_dma_sync_single_for_device,
198 .sync_sg_for_cpu = arm_dma_sync_sg_for_cpu,
199 .sync_sg_for_device = arm_dma_sync_sg_for_device,
200 .mapping_error = arm_dma_mapping_error,
201 .dma_supported = arm_dma_supported,
203 EXPORT_SYMBOL(arm_dma_ops);
205 static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
206 dma_addr_t *handle, gfp_t gfp, unsigned long attrs);
207 static void arm_coherent_dma_free(struct device *dev, size_t size, void *cpu_addr,
208 dma_addr_t handle, unsigned long attrs);
209 static int arm_coherent_dma_mmap(struct device *dev, struct vm_area_struct *vma,
210 void *cpu_addr, dma_addr_t dma_addr, size_t size,
211 unsigned long attrs);
213 const struct dma_map_ops arm_coherent_dma_ops = {
214 .alloc = arm_coherent_dma_alloc,
215 .free = arm_coherent_dma_free,
216 .mmap = arm_coherent_dma_mmap,
217 .get_sgtable = arm_dma_get_sgtable,
218 .map_page = arm_coherent_dma_map_page,
219 .map_sg = arm_dma_map_sg,
220 .mapping_error = arm_dma_mapping_error,
221 .dma_supported = arm_dma_supported,
223 EXPORT_SYMBOL(arm_coherent_dma_ops);
225 static int __dma_supported(struct device *dev, u64 mask, bool warn)
227 unsigned long max_dma_pfn;
230 * If the mask allows for more memory than we can address,
231 * and we actually have that much memory, then we must
232 * indicate that DMA to this device is not supported.
234 if (sizeof(mask) != sizeof(dma_addr_t) &&
235 mask > (dma_addr_t)~0 &&
236 dma_to_pfn(dev, ~0) < max_pfn - 1) {
238 dev_warn(dev, "Coherent DMA mask %#llx is larger than dma_addr_t allows\n",
240 dev_warn(dev, "Driver did not use or check the return value from dma_set_coherent_mask()?\n");
245 max_dma_pfn = min(max_pfn, arm_dma_pfn_limit);
248 * Translate the device's DMA mask to a PFN limit. This
249 * PFN number includes the page which we can DMA to.
251 if (dma_to_pfn(dev, mask) < max_dma_pfn) {
253 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 dma_to_pfn(dev, 0), dma_to_pfn(dev, mask) + 1,
263 static u64 get_coherent_dma_mask(struct device *dev)
265 u64 mask = (u64)DMA_BIT_MASK(32);
268 mask = dev->coherent_dma_mask;
271 * Sanity check the DMA mask - it must be non-zero, and
272 * must be able to be satisfied by a DMA allocation.
275 dev_warn(dev, "coherent DMA mask is unset\n");
279 if (!__dma_supported(dev, mask, true))
286 static void __dma_clear_buffer(struct page *page, size_t size, int coherent_flag)
289 * Ensure that the allocated pages are zeroed, and that any data
290 * lurking in the kernel direct-mapped region is invalidated.
292 if (PageHighMem(page)) {
293 phys_addr_t base = __pfn_to_phys(page_to_pfn(page));
294 phys_addr_t end = base + size;
296 void *ptr = kmap_atomic(page);
297 memset(ptr, 0, PAGE_SIZE);
298 if (coherent_flag != COHERENT)
299 dmac_flush_range(ptr, ptr + PAGE_SIZE);
304 if (coherent_flag != COHERENT)
305 outer_flush_range(base, end);
307 void *ptr = page_address(page);
308 memset(ptr, 0, size);
309 if (coherent_flag != COHERENT) {
310 dmac_flush_range(ptr, ptr + size);
311 outer_flush_range(__pa(ptr), __pa(ptr) + size);
317 * Allocate a DMA buffer for 'dev' of size 'size' using the
318 * specified gfp mask. Note that 'size' must be page aligned.
320 static struct page *__dma_alloc_buffer(struct device *dev, size_t size,
321 gfp_t gfp, int coherent_flag)
323 unsigned long order = get_order(size);
324 struct page *page, *p, *e;
326 page = alloc_pages(gfp, order);
331 * Now split the huge page and free the excess pages
333 split_page(page, order);
334 for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
337 __dma_clear_buffer(page, size, coherent_flag);
343 * Free a DMA buffer. 'size' must be page aligned.
345 static void __dma_free_buffer(struct page *page, size_t size)
347 struct page *e = page + (size >> PAGE_SHIFT);
355 static void *__alloc_from_contiguous(struct device *dev, size_t size,
356 pgprot_t prot, struct page **ret_page,
357 const void *caller, bool want_vaddr,
358 int coherent_flag, gfp_t gfp);
360 static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
361 pgprot_t prot, struct page **ret_page,
362 const void *caller, bool want_vaddr);
365 __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot,
369 * DMA allocation can be mapped to user space, so lets
370 * set VM_USERMAP flags too.
372 return dma_common_contiguous_remap(page, size,
373 VM_ARM_DMA_CONSISTENT | VM_USERMAP,
377 static void __dma_free_remap(void *cpu_addr, size_t size)
379 dma_common_free_remap(cpu_addr, size,
380 VM_ARM_DMA_CONSISTENT | VM_USERMAP);
383 #define DEFAULT_DMA_COHERENT_POOL_SIZE SZ_256K
384 static struct gen_pool *atomic_pool __ro_after_init;
386 static size_t atomic_pool_size __initdata = DEFAULT_DMA_COHERENT_POOL_SIZE;
388 static int __init early_coherent_pool(char *p)
390 atomic_pool_size = memparse(p, &p);
393 early_param("coherent_pool", early_coherent_pool);
396 * Initialise the coherent pool for atomic allocations.
398 static int __init atomic_pool_init(void)
400 pgprot_t prot = pgprot_dmacoherent(PAGE_KERNEL);
401 gfp_t gfp = GFP_KERNEL | GFP_DMA;
405 atomic_pool = gen_pool_create(PAGE_SHIFT, -1);
409 * The atomic pool is only used for non-coherent allocations
410 * so we must pass NORMAL for coherent_flag.
412 if (dev_get_cma_area(NULL))
413 ptr = __alloc_from_contiguous(NULL, atomic_pool_size, prot,
414 &page, atomic_pool_init, true, NORMAL,
417 ptr = __alloc_remap_buffer(NULL, atomic_pool_size, gfp, prot,
418 &page, atomic_pool_init, true);
422 ret = gen_pool_add_virt(atomic_pool, (unsigned long)ptr,
424 atomic_pool_size, -1);
426 goto destroy_genpool;
428 gen_pool_set_algo(atomic_pool,
429 gen_pool_first_fit_order_align,
431 pr_info("DMA: preallocated %zu KiB pool for atomic coherent allocations\n",
432 atomic_pool_size / 1024);
437 gen_pool_destroy(atomic_pool);
440 pr_err("DMA: failed to allocate %zu KiB pool for atomic coherent allocation\n",
441 atomic_pool_size / 1024);
445 * CMA is activated by core_initcall, so we must be called after it.
447 postcore_initcall(atomic_pool_init);
449 struct dma_contig_early_reserve {
454 static struct dma_contig_early_reserve dma_mmu_remap[MAX_CMA_AREAS] __initdata;
456 static int dma_mmu_remap_num __initdata;
458 void __init dma_contiguous_early_fixup(phys_addr_t base, unsigned long size)
460 dma_mmu_remap[dma_mmu_remap_num].base = base;
461 dma_mmu_remap[dma_mmu_remap_num].size = size;
465 void __init dma_contiguous_remap(void)
468 for (i = 0; i < dma_mmu_remap_num; i++) {
469 phys_addr_t start = dma_mmu_remap[i].base;
470 phys_addr_t end = start + dma_mmu_remap[i].size;
474 if (end > arm_lowmem_limit)
475 end = arm_lowmem_limit;
479 map.pfn = __phys_to_pfn(start);
480 map.virtual = __phys_to_virt(start);
481 map.length = end - start;
482 map.type = MT_MEMORY_DMA_READY;
485 * Clear previous low-memory mapping to ensure that the
486 * TLB does not see any conflicting entries, then flush
487 * the TLB of the old entries before creating new mappings.
489 * This ensures that any speculatively loaded TLB entries
490 * (even though they may be rare) can not cause any problems,
491 * and ensures that this code is architecturally compliant.
493 for (addr = __phys_to_virt(start); addr < __phys_to_virt(end);
495 pmd_clear(pmd_off_k(addr));
497 flush_tlb_kernel_range(__phys_to_virt(start),
498 __phys_to_virt(end));
500 iotable_init(&map, 1);
504 static int __dma_update_pte(pte_t *pte, pgtable_t token, unsigned long addr,
507 struct page *page = virt_to_page(addr);
508 pgprot_t prot = *(pgprot_t *)data;
510 set_pte_ext(pte, mk_pte(page, prot), 0);
514 static void __dma_remap(struct page *page, size_t size, pgprot_t prot)
516 unsigned long start = (unsigned long) page_address(page);
517 unsigned end = start + size;
519 apply_to_page_range(&init_mm, start, size, __dma_update_pte, &prot);
520 flush_tlb_kernel_range(start, end);
523 static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
524 pgprot_t prot, struct page **ret_page,
525 const void *caller, bool want_vaddr)
530 * __alloc_remap_buffer is only called when the device is
533 page = __dma_alloc_buffer(dev, size, gfp, NORMAL);
539 ptr = __dma_alloc_remap(page, size, gfp, prot, caller);
541 __dma_free_buffer(page, size);
550 static void *__alloc_from_pool(size_t size, struct page **ret_page)
556 WARN(1, "coherent pool not initialised!\n");
560 val = gen_pool_alloc(atomic_pool, size);
562 phys_addr_t phys = gen_pool_virt_to_phys(atomic_pool, val);
564 *ret_page = phys_to_page(phys);
571 static bool __in_atomic_pool(void *start, size_t size)
573 return addr_in_gen_pool(atomic_pool, (unsigned long)start, size);
576 static int __free_from_pool(void *start, size_t size)
578 if (!__in_atomic_pool(start, size))
581 gen_pool_free(atomic_pool, (unsigned long)start, size);
586 static void *__alloc_from_contiguous(struct device *dev, size_t size,
587 pgprot_t prot, struct page **ret_page,
588 const void *caller, bool want_vaddr,
589 int coherent_flag, gfp_t gfp)
591 unsigned long order = get_order(size);
592 size_t count = size >> PAGE_SHIFT;
596 page = dma_alloc_from_contiguous(dev, count, order, gfp & __GFP_NOWARN);
600 __dma_clear_buffer(page, size, coherent_flag);
605 if (PageHighMem(page)) {
606 ptr = __dma_alloc_remap(page, size, GFP_KERNEL, prot, caller);
608 dma_release_from_contiguous(dev, page, count);
612 __dma_remap(page, size, prot);
613 ptr = page_address(page);
621 static void __free_from_contiguous(struct device *dev, struct page *page,
622 void *cpu_addr, size_t size, bool want_vaddr)
625 if (PageHighMem(page))
626 __dma_free_remap(cpu_addr, size);
628 __dma_remap(page, size, PAGE_KERNEL);
630 dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT);
633 static inline pgprot_t __get_dma_pgprot(unsigned long attrs, pgprot_t prot)
635 prot = (attrs & DMA_ATTR_WRITE_COMBINE) ?
636 pgprot_writecombine(prot) :
637 pgprot_dmacoherent(prot);
641 static void *__alloc_simple_buffer(struct device *dev, size_t size, gfp_t gfp,
642 struct page **ret_page)
645 /* __alloc_simple_buffer is only called when the device is coherent */
646 page = __dma_alloc_buffer(dev, size, gfp, COHERENT);
651 return page_address(page);
654 static void *simple_allocator_alloc(struct arm_dma_alloc_args *args,
655 struct page **ret_page)
657 return __alloc_simple_buffer(args->dev, args->size, args->gfp,
661 static void simple_allocator_free(struct arm_dma_free_args *args)
663 __dma_free_buffer(args->page, args->size);
666 static struct arm_dma_allocator simple_allocator = {
667 .alloc = simple_allocator_alloc,
668 .free = simple_allocator_free,
671 static void *cma_allocator_alloc(struct arm_dma_alloc_args *args,
672 struct page **ret_page)
674 return __alloc_from_contiguous(args->dev, args->size, args->prot,
675 ret_page, args->caller,
676 args->want_vaddr, args->coherent_flag,
680 static void cma_allocator_free(struct arm_dma_free_args *args)
682 __free_from_contiguous(args->dev, args->page, args->cpu_addr,
683 args->size, args->want_vaddr);
686 static struct arm_dma_allocator cma_allocator = {
687 .alloc = cma_allocator_alloc,
688 .free = cma_allocator_free,
691 static void *pool_allocator_alloc(struct arm_dma_alloc_args *args,
692 struct page **ret_page)
694 return __alloc_from_pool(args->size, ret_page);
697 static void pool_allocator_free(struct arm_dma_free_args *args)
699 __free_from_pool(args->cpu_addr, args->size);
702 static struct arm_dma_allocator pool_allocator = {
703 .alloc = pool_allocator_alloc,
704 .free = pool_allocator_free,
707 static void *remap_allocator_alloc(struct arm_dma_alloc_args *args,
708 struct page **ret_page)
710 return __alloc_remap_buffer(args->dev, args->size, args->gfp,
711 args->prot, ret_page, args->caller,
715 static void remap_allocator_free(struct arm_dma_free_args *args)
717 if (args->want_vaddr)
718 __dma_free_remap(args->cpu_addr, args->size);
720 __dma_free_buffer(args->page, args->size);
723 static struct arm_dma_allocator remap_allocator = {
724 .alloc = remap_allocator_alloc,
725 .free = remap_allocator_free,
728 static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
729 gfp_t gfp, pgprot_t prot, bool is_coherent,
730 unsigned long attrs, const void *caller)
732 u64 mask = get_coherent_dma_mask(dev);
733 struct page *page = NULL;
735 bool allowblock, cma;
736 struct arm_dma_buffer *buf;
737 struct arm_dma_alloc_args args = {
739 .size = PAGE_ALIGN(size),
743 .want_vaddr = ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) == 0),
744 .coherent_flag = is_coherent ? COHERENT : NORMAL,
747 #ifdef CONFIG_DMA_API_DEBUG
748 u64 limit = (mask + 1) & ~mask;
749 if (limit && size >= limit) {
750 dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n",
759 buf = kzalloc(sizeof(*buf),
760 gfp & ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM));
764 if (mask < 0xffffffffULL)
768 * Following is a work-around (a.k.a. hack) to prevent pages
769 * with __GFP_COMP being passed to split_page() which cannot
770 * handle them. The real problem is that this flag probably
771 * should be 0 on ARM as it is not supported on this
772 * platform; see CONFIG_HUGETLBFS.
774 gfp &= ~(__GFP_COMP);
777 *handle = ARM_MAPPING_ERROR;
778 allowblock = gfpflags_allow_blocking(gfp);
779 cma = allowblock ? dev_get_cma_area(dev) : false;
782 buf->allocator = &cma_allocator;
783 else if (is_coherent)
784 buf->allocator = &simple_allocator;
786 buf->allocator = &remap_allocator;
788 buf->allocator = &pool_allocator;
790 addr = buf->allocator->alloc(&args, &page);
795 *handle = pfn_to_dma(dev, page_to_pfn(page));
796 buf->virt = args.want_vaddr ? addr : page;
798 spin_lock_irqsave(&arm_dma_bufs_lock, flags);
799 list_add(&buf->list, &arm_dma_bufs);
800 spin_unlock_irqrestore(&arm_dma_bufs_lock, flags);
805 return args.want_vaddr ? addr : page;
809 * Allocate DMA-coherent memory space and return both the kernel remapped
810 * virtual and bus address for that space.
812 void *arm_dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
813 gfp_t gfp, unsigned long attrs)
815 pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
817 return __dma_alloc(dev, size, handle, gfp, prot, false,
818 attrs, __builtin_return_address(0));
821 static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
822 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
824 return __dma_alloc(dev, size, handle, gfp, PAGE_KERNEL, true,
825 attrs, __builtin_return_address(0));
828 static int __arm_dma_mmap(struct device *dev, struct vm_area_struct *vma,
829 void *cpu_addr, dma_addr_t dma_addr, size_t size,
833 unsigned long nr_vma_pages = vma_pages(vma);
834 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
835 unsigned long pfn = dma_to_pfn(dev, dma_addr);
836 unsigned long off = vma->vm_pgoff;
838 if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
841 if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) {
842 ret = remap_pfn_range(vma, vma->vm_start,
844 vma->vm_end - vma->vm_start,
852 * Create userspace mapping for the DMA-coherent memory.
854 static int arm_coherent_dma_mmap(struct device *dev, struct vm_area_struct *vma,
855 void *cpu_addr, dma_addr_t dma_addr, size_t size,
858 return __arm_dma_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
861 int arm_dma_mmap(struct device *dev, struct vm_area_struct *vma,
862 void *cpu_addr, dma_addr_t dma_addr, size_t size,
865 vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot);
866 return __arm_dma_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
870 * Free a buffer as defined by the above mapping.
872 static void __arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
873 dma_addr_t handle, unsigned long attrs,
876 struct page *page = pfn_to_page(dma_to_pfn(dev, handle));
877 struct arm_dma_buffer *buf;
878 struct arm_dma_free_args args = {
880 .size = PAGE_ALIGN(size),
881 .cpu_addr = cpu_addr,
883 .want_vaddr = ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) == 0),
886 buf = arm_dma_buffer_find(cpu_addr);
887 if (WARN(!buf, "Freeing invalid buffer %p\n", cpu_addr))
890 buf->allocator->free(&args);
894 void arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
895 dma_addr_t handle, unsigned long attrs)
897 __arm_dma_free(dev, size, cpu_addr, handle, attrs, false);
900 static void arm_coherent_dma_free(struct device *dev, size_t size, void *cpu_addr,
901 dma_addr_t handle, unsigned long attrs)
903 __arm_dma_free(dev, size, cpu_addr, handle, attrs, true);
907 * The whole dma_get_sgtable() idea is fundamentally unsafe - it seems
908 * that the intention is to allow exporting memory allocated via the
909 * coherent DMA APIs through the dma_buf API, which only accepts a
910 * scattertable. This presents a couple of problems:
911 * 1. Not all memory allocated via the coherent DMA APIs is backed by
913 * 2. Passing coherent DMA memory into the streaming APIs is not allowed
914 * as we will try to flush the memory through a different alias to that
915 * actually being used (and the flushes are redundant.)
917 int arm_dma_get_sgtable(struct device *dev, struct sg_table *sgt,
918 void *cpu_addr, dma_addr_t handle, size_t size,
921 unsigned long pfn = dma_to_pfn(dev, handle);
925 /* If the PFN is not valid, we do not have a struct page */
929 page = pfn_to_page(pfn);
931 ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
935 sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
939 static void dma_cache_maint_page(struct page *page, unsigned long offset,
940 size_t size, enum dma_data_direction dir,
941 void (*op)(const void *, size_t, int))
946 pfn = page_to_pfn(page) + offset / PAGE_SIZE;
950 * A single sg entry may refer to multiple physically contiguous
951 * pages. But we still need to process highmem pages individually.
952 * If highmem is not configured then the bulk of this loop gets
959 page = pfn_to_page(pfn);
961 if (PageHighMem(page)) {
962 if (len + offset > PAGE_SIZE)
963 len = PAGE_SIZE - offset;
965 if (cache_is_vipt_nonaliasing()) {
966 vaddr = kmap_atomic(page);
967 op(vaddr + offset, len, dir);
968 kunmap_atomic(vaddr);
970 vaddr = kmap_high_get(page);
972 op(vaddr + offset, len, dir);
977 vaddr = page_address(page) + offset;
987 * Make an area consistent for devices.
988 * Note: Drivers should NOT use this function directly, as it will break
989 * platforms with CONFIG_DMABOUNCE.
990 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
992 static void __dma_page_cpu_to_dev(struct page *page, unsigned long off,
993 size_t size, enum dma_data_direction dir)
997 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
999 paddr = page_to_phys(page) + off;
1000 if (dir == DMA_FROM_DEVICE) {
1001 outer_inv_range(paddr, paddr + size);
1003 outer_clean_range(paddr, paddr + size);
1005 /* FIXME: non-speculating: flush on bidirectional mappings? */
1008 static void __dma_page_dev_to_cpu(struct page *page, unsigned long off,
1009 size_t size, enum dma_data_direction dir)
1011 phys_addr_t paddr = page_to_phys(page) + off;
1013 /* FIXME: non-speculating: not required */
1014 /* in any case, don't bother invalidating if DMA to device */
1015 if (dir != DMA_TO_DEVICE) {
1016 outer_inv_range(paddr, paddr + size);
1018 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
1022 * Mark the D-cache clean for these pages to avoid extra flushing.
1024 if (dir != DMA_TO_DEVICE && size >= PAGE_SIZE) {
1028 pfn = page_to_pfn(page) + off / PAGE_SIZE;
1032 left -= PAGE_SIZE - off;
1034 while (left >= PAGE_SIZE) {
1035 page = pfn_to_page(pfn++);
1036 set_bit(PG_dcache_clean, &page->flags);
1043 * arm_dma_map_sg - map a set of SG buffers for streaming mode DMA
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
1047 * @dir: DMA transfer direction
1049 * Map a set of buffers described by scatterlist in streaming mode for DMA.
1050 * This is the scatter-gather version of the dma_map_single interface.
1051 * Here the scatter gather list elements are each tagged with the
1052 * appropriate dma address and length. They are obtained via
1053 * sg_dma_{address,length}.
1055 * Device ownership issues as mentioned for dma_map_single are the same
1058 int arm_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
1059 enum dma_data_direction dir, unsigned long attrs)
1061 const struct dma_map_ops *ops = get_dma_ops(dev);
1062 struct scatterlist *s;
1065 for_each_sg(sg, s, nents, i) {
1066 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1067 s->dma_length = s->length;
1069 s->dma_address = ops->map_page(dev, sg_page(s), s->offset,
1070 s->length, dir, attrs);
1071 if (dma_mapping_error(dev, s->dma_address))
1077 for_each_sg(sg, s, i, j)
1078 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
1083 * arm_dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
1084 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
1085 * @sg: list of buffers
1086 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
1087 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1089 * Unmap a set of streaming mode DMA translations. Again, CPU access
1090 * rules concerning calls here are the same as for dma_unmap_single().
1092 void arm_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
1093 enum dma_data_direction dir, unsigned long attrs)
1095 const struct dma_map_ops *ops = get_dma_ops(dev);
1096 struct scatterlist *s;
1100 for_each_sg(sg, s, nents, i)
1101 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
1105 * arm_dma_sync_sg_for_cpu
1106 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
1107 * @sg: list of buffers
1108 * @nents: number of buffers to map (returned from dma_map_sg)
1109 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1111 void arm_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1112 int nents, enum dma_data_direction dir)
1114 const struct dma_map_ops *ops = get_dma_ops(dev);
1115 struct scatterlist *s;
1118 for_each_sg(sg, s, nents, i)
1119 ops->sync_single_for_cpu(dev, sg_dma_address(s), s->length,
1124 * arm_dma_sync_sg_for_device
1125 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
1126 * @sg: list of buffers
1127 * @nents: number of buffers to map (returned from dma_map_sg)
1128 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1130 void arm_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1131 int nents, enum dma_data_direction dir)
1133 const struct dma_map_ops *ops = get_dma_ops(dev);
1134 struct scatterlist *s;
1137 for_each_sg(sg, s, nents, i)
1138 ops->sync_single_for_device(dev, sg_dma_address(s), s->length,
1143 * Return whether the given device DMA address mask can be supported
1144 * properly. For example, if your device can only drive the low 24-bits
1145 * during bus mastering, then you would pass 0x00ffffff as the mask
1148 int arm_dma_supported(struct device *dev, u64 mask)
1150 return __dma_supported(dev, mask, false);
1153 static const struct dma_map_ops *arm_get_dma_map_ops(bool coherent)
1155 return coherent ? &arm_coherent_dma_ops : &arm_dma_ops;
1158 #ifdef CONFIG_ARM_DMA_USE_IOMMU
1160 static int __dma_info_to_prot(enum dma_data_direction dir, unsigned long attrs)
1164 if (attrs & DMA_ATTR_PRIVILEGED)
1168 case DMA_BIDIRECTIONAL:
1169 return prot | IOMMU_READ | IOMMU_WRITE;
1171 return prot | IOMMU_READ;
1172 case DMA_FROM_DEVICE:
1173 return prot | IOMMU_WRITE;
1181 static int extend_iommu_mapping(struct dma_iommu_mapping *mapping);
1183 static inline dma_addr_t __alloc_iova(struct dma_iommu_mapping *mapping,
1186 unsigned int order = get_order(size);
1187 unsigned int align = 0;
1188 unsigned int count, start;
1189 size_t mapping_size = mapping->bits << PAGE_SHIFT;
1190 unsigned long flags;
1194 if (order > CONFIG_ARM_DMA_IOMMU_ALIGNMENT)
1195 order = CONFIG_ARM_DMA_IOMMU_ALIGNMENT;
1197 count = PAGE_ALIGN(size) >> PAGE_SHIFT;
1198 align = (1 << order) - 1;
1200 spin_lock_irqsave(&mapping->lock, flags);
1201 for (i = 0; i < mapping->nr_bitmaps; i++) {
1202 start = bitmap_find_next_zero_area(mapping->bitmaps[i],
1203 mapping->bits, 0, count, align);
1205 if (start > mapping->bits)
1208 bitmap_set(mapping->bitmaps[i], start, count);
1213 * No unused range found. Try to extend the existing mapping
1214 * and perform a second attempt to reserve an IO virtual
1215 * address range of size bytes.
1217 if (i == mapping->nr_bitmaps) {
1218 if (extend_iommu_mapping(mapping)) {
1219 spin_unlock_irqrestore(&mapping->lock, flags);
1220 return ARM_MAPPING_ERROR;
1223 start = bitmap_find_next_zero_area(mapping->bitmaps[i],
1224 mapping->bits, 0, count, align);
1226 if (start > mapping->bits) {
1227 spin_unlock_irqrestore(&mapping->lock, flags);
1228 return ARM_MAPPING_ERROR;
1231 bitmap_set(mapping->bitmaps[i], start, count);
1233 spin_unlock_irqrestore(&mapping->lock, flags);
1235 iova = mapping->base + (mapping_size * i);
1236 iova += start << PAGE_SHIFT;
1241 static inline void __free_iova(struct dma_iommu_mapping *mapping,
1242 dma_addr_t addr, size_t size)
1244 unsigned int start, count;
1245 size_t mapping_size = mapping->bits << PAGE_SHIFT;
1246 unsigned long flags;
1247 dma_addr_t bitmap_base;
1253 bitmap_index = (u32) (addr - mapping->base) / (u32) mapping_size;
1254 BUG_ON(addr < mapping->base || bitmap_index > mapping->extensions);
1256 bitmap_base = mapping->base + mapping_size * bitmap_index;
1258 start = (addr - bitmap_base) >> PAGE_SHIFT;
1260 if (addr + size > bitmap_base + mapping_size) {
1262 * The address range to be freed reaches into the iova
1263 * range of the next bitmap. This should not happen as
1264 * we don't allow this in __alloc_iova (at the
1269 count = size >> PAGE_SHIFT;
1271 spin_lock_irqsave(&mapping->lock, flags);
1272 bitmap_clear(mapping->bitmaps[bitmap_index], start, count);
1273 spin_unlock_irqrestore(&mapping->lock, flags);
1276 /* We'll try 2M, 1M, 64K, and finally 4K; array must end with 0! */
1277 static const int iommu_order_array[] = { 9, 8, 4, 0 };
1279 static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
1280 gfp_t gfp, unsigned long attrs,
1283 struct page **pages;
1284 int count = size >> PAGE_SHIFT;
1285 int array_size = count * sizeof(struct page *);
1289 if (array_size <= PAGE_SIZE)
1290 pages = kzalloc(array_size, GFP_KERNEL);
1292 pages = vzalloc(array_size);
1296 if (attrs & DMA_ATTR_FORCE_CONTIGUOUS)
1298 unsigned long order = get_order(size);
1301 page = dma_alloc_from_contiguous(dev, count, order,
1302 gfp & __GFP_NOWARN);
1306 __dma_clear_buffer(page, size, coherent_flag);
1308 for (i = 0; i < count; i++)
1309 pages[i] = page + i;
1314 /* Go straight to 4K chunks if caller says it's OK. */
1315 if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)
1316 order_idx = ARRAY_SIZE(iommu_order_array) - 1;
1319 * IOMMU can map any pages, so himem can also be used here
1321 gfp |= __GFP_NOWARN | __GFP_HIGHMEM;
1326 order = iommu_order_array[order_idx];
1328 /* Drop down when we get small */
1329 if (__fls(count) < order) {
1335 /* See if it's easy to allocate a high-order chunk */
1336 pages[i] = alloc_pages(gfp | __GFP_NORETRY, order);
1338 /* Go down a notch at first sign of pressure */
1344 pages[i] = alloc_pages(gfp, 0);
1350 split_page(pages[i], order);
1353 pages[i + j] = pages[i] + j;
1356 __dma_clear_buffer(pages[i], PAGE_SIZE << order, coherent_flag);
1358 count -= 1 << order;
1365 __free_pages(pages[i], 0);
1370 static int __iommu_free_buffer(struct device *dev, struct page **pages,
1371 size_t size, unsigned long attrs)
1373 int count = size >> PAGE_SHIFT;
1376 if (attrs & DMA_ATTR_FORCE_CONTIGUOUS) {
1377 dma_release_from_contiguous(dev, pages[0], count);
1379 for (i = 0; i < count; i++)
1381 __free_pages(pages[i], 0);
1389 * Create a CPU mapping for a specified pages
1392 __iommu_alloc_remap(struct page **pages, size_t size, gfp_t gfp, pgprot_t prot,
1395 return dma_common_pages_remap(pages, size,
1396 VM_ARM_DMA_CONSISTENT | VM_USERMAP, prot, caller);
1400 * Create a mapping in device IO address space for specified pages
1403 __iommu_create_mapping(struct device *dev, struct page **pages, size_t size,
1404 unsigned long attrs)
1406 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1407 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
1408 dma_addr_t dma_addr, iova;
1411 dma_addr = __alloc_iova(mapping, size);
1412 if (dma_addr == ARM_MAPPING_ERROR)
1416 for (i = 0; i < count; ) {
1419 unsigned int next_pfn = page_to_pfn(pages[i]) + 1;
1420 phys_addr_t phys = page_to_phys(pages[i]);
1421 unsigned int len, j;
1423 for (j = i + 1; j < count; j++, next_pfn++)
1424 if (page_to_pfn(pages[j]) != next_pfn)
1427 len = (j - i) << PAGE_SHIFT;
1428 ret = iommu_map(mapping->domain, iova, phys, len,
1429 __dma_info_to_prot(DMA_BIDIRECTIONAL, attrs));
1437 iommu_unmap(mapping->domain, dma_addr, iova-dma_addr);
1438 __free_iova(mapping, dma_addr, size);
1439 return ARM_MAPPING_ERROR;
1442 static int __iommu_remove_mapping(struct device *dev, dma_addr_t iova, size_t size)
1444 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1447 * add optional in-page offset from iova to size and align
1448 * result to page size
1450 size = PAGE_ALIGN((iova & ~PAGE_MASK) + size);
1453 iommu_unmap(mapping->domain, iova, size);
1454 __free_iova(mapping, iova, size);
1458 static struct page **__atomic_get_pages(void *addr)
1463 phys = gen_pool_virt_to_phys(atomic_pool, (unsigned long)addr);
1464 page = phys_to_page(phys);
1466 return (struct page **)page;
1469 static struct page **__iommu_get_pages(void *cpu_addr, unsigned long attrs)
1471 struct vm_struct *area;
1473 if (__in_atomic_pool(cpu_addr, PAGE_SIZE))
1474 return __atomic_get_pages(cpu_addr);
1476 if (attrs & DMA_ATTR_NO_KERNEL_MAPPING)
1479 area = find_vm_area(cpu_addr);
1480 if (area && (area->flags & VM_ARM_DMA_CONSISTENT))
1485 static void *__iommu_alloc_simple(struct device *dev, size_t size, gfp_t gfp,
1486 dma_addr_t *handle, int coherent_flag,
1487 unsigned long attrs)
1492 if (coherent_flag == COHERENT)
1493 addr = __alloc_simple_buffer(dev, size, gfp, &page);
1495 addr = __alloc_from_pool(size, &page);
1499 *handle = __iommu_create_mapping(dev, &page, size, attrs);
1500 if (*handle == ARM_MAPPING_ERROR)
1506 __free_from_pool(addr, size);
1510 static void __iommu_free_atomic(struct device *dev, void *cpu_addr,
1511 dma_addr_t handle, size_t size, int coherent_flag)
1513 __iommu_remove_mapping(dev, handle, size);
1514 if (coherent_flag == COHERENT)
1515 __dma_free_buffer(virt_to_page(cpu_addr), size);
1517 __free_from_pool(cpu_addr, size);
1520 static void *__arm_iommu_alloc_attrs(struct device *dev, size_t size,
1521 dma_addr_t *handle, gfp_t gfp, unsigned long attrs,
1524 pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
1525 struct page **pages;
1528 *handle = ARM_MAPPING_ERROR;
1529 size = PAGE_ALIGN(size);
1531 if (coherent_flag == COHERENT || !gfpflags_allow_blocking(gfp))
1532 return __iommu_alloc_simple(dev, size, gfp, handle,
1533 coherent_flag, attrs);
1536 * Following is a work-around (a.k.a. hack) to prevent pages
1537 * with __GFP_COMP being passed to split_page() which cannot
1538 * handle them. The real problem is that this flag probably
1539 * should be 0 on ARM as it is not supported on this
1540 * platform; see CONFIG_HUGETLBFS.
1542 gfp &= ~(__GFP_COMP);
1544 pages = __iommu_alloc_buffer(dev, size, gfp, attrs, coherent_flag);
1548 *handle = __iommu_create_mapping(dev, pages, size, attrs);
1549 if (*handle == ARM_MAPPING_ERROR)
1552 if (attrs & DMA_ATTR_NO_KERNEL_MAPPING)
1555 addr = __iommu_alloc_remap(pages, size, gfp, prot,
1556 __builtin_return_address(0));
1563 __iommu_remove_mapping(dev, *handle, size);
1565 __iommu_free_buffer(dev, pages, size, attrs);
1569 static void *arm_iommu_alloc_attrs(struct device *dev, size_t size,
1570 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1572 return __arm_iommu_alloc_attrs(dev, size, handle, gfp, attrs, NORMAL);
1575 static void *arm_coherent_iommu_alloc_attrs(struct device *dev, size_t size,
1576 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1578 return __arm_iommu_alloc_attrs(dev, size, handle, gfp, attrs, COHERENT);
1581 static int __arm_iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
1582 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1583 unsigned long attrs)
1585 unsigned long uaddr = vma->vm_start;
1586 unsigned long usize = vma->vm_end - vma->vm_start;
1587 struct page **pages = __iommu_get_pages(cpu_addr, attrs);
1588 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
1589 unsigned long off = vma->vm_pgoff;
1594 if (off >= nr_pages || (usize >> PAGE_SHIFT) > nr_pages - off)
1600 int ret = vm_insert_page(vma, uaddr, *pages++);
1602 pr_err("Remapping memory failed: %d\n", ret);
1607 } while (usize > 0);
1611 static int arm_iommu_mmap_attrs(struct device *dev,
1612 struct vm_area_struct *vma, void *cpu_addr,
1613 dma_addr_t dma_addr, size_t size, unsigned long attrs)
1615 vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot);
1617 return __arm_iommu_mmap_attrs(dev, vma, cpu_addr, dma_addr, size, attrs);
1620 static int arm_coherent_iommu_mmap_attrs(struct device *dev,
1621 struct vm_area_struct *vma, void *cpu_addr,
1622 dma_addr_t dma_addr, size_t size, unsigned long attrs)
1624 return __arm_iommu_mmap_attrs(dev, vma, cpu_addr, dma_addr, size, attrs);
1628 * free a page as defined by the above mapping.
1629 * Must not be called with IRQs disabled.
1631 void __arm_iommu_free_attrs(struct device *dev, size_t size, void *cpu_addr,
1632 dma_addr_t handle, unsigned long attrs, int coherent_flag)
1634 struct page **pages;
1635 size = PAGE_ALIGN(size);
1637 if (coherent_flag == COHERENT || __in_atomic_pool(cpu_addr, size)) {
1638 __iommu_free_atomic(dev, cpu_addr, handle, size, coherent_flag);
1642 pages = __iommu_get_pages(cpu_addr, attrs);
1644 WARN(1, "trying to free invalid coherent area: %p\n", cpu_addr);
1648 if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) == 0) {
1649 dma_common_free_remap(cpu_addr, size,
1650 VM_ARM_DMA_CONSISTENT | VM_USERMAP);
1653 __iommu_remove_mapping(dev, handle, size);
1654 __iommu_free_buffer(dev, pages, size, attrs);
1657 void arm_iommu_free_attrs(struct device *dev, size_t size,
1658 void *cpu_addr, dma_addr_t handle, unsigned long attrs)
1660 __arm_iommu_free_attrs(dev, size, cpu_addr, handle, attrs, NORMAL);
1663 void arm_coherent_iommu_free_attrs(struct device *dev, size_t size,
1664 void *cpu_addr, dma_addr_t handle, unsigned long attrs)
1666 __arm_iommu_free_attrs(dev, size, cpu_addr, handle, attrs, COHERENT);
1669 static int arm_iommu_get_sgtable(struct device *dev, struct sg_table *sgt,
1670 void *cpu_addr, dma_addr_t dma_addr,
1671 size_t size, unsigned long attrs)
1673 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
1674 struct page **pages = __iommu_get_pages(cpu_addr, attrs);
1679 return sg_alloc_table_from_pages(sgt, pages, count, 0, size,
1684 * Map a part of the scatter-gather list into contiguous io address space
1686 static int __map_sg_chunk(struct device *dev, struct scatterlist *sg,
1687 size_t size, dma_addr_t *handle,
1688 enum dma_data_direction dir, unsigned long attrs,
1691 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1692 dma_addr_t iova, iova_base;
1695 struct scatterlist *s;
1698 size = PAGE_ALIGN(size);
1699 *handle = ARM_MAPPING_ERROR;
1701 iova_base = iova = __alloc_iova(mapping, size);
1702 if (iova == ARM_MAPPING_ERROR)
1705 for (count = 0, s = sg; count < (size >> PAGE_SHIFT); s = sg_next(s)) {
1706 phys_addr_t phys = page_to_phys(sg_page(s));
1707 unsigned int len = PAGE_ALIGN(s->offset + s->length);
1709 if (!is_coherent && (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1710 __dma_page_cpu_to_dev(sg_page(s), s->offset, s->length, dir);
1712 prot = __dma_info_to_prot(dir, attrs);
1714 ret = iommu_map(mapping->domain, iova, phys, len, prot);
1717 count += len >> PAGE_SHIFT;
1720 *handle = iova_base;
1724 iommu_unmap(mapping->domain, iova_base, count * PAGE_SIZE);
1725 __free_iova(mapping, iova_base, size);
1729 static int __iommu_map_sg(struct device *dev, struct scatterlist *sg, int nents,
1730 enum dma_data_direction dir, unsigned long attrs,
1733 struct scatterlist *s = sg, *dma = sg, *start = sg;
1735 unsigned int offset = s->offset;
1736 unsigned int size = s->offset + s->length;
1737 unsigned int max = dma_get_max_seg_size(dev);
1739 for (i = 1; i < nents; i++) {
1742 s->dma_address = ARM_MAPPING_ERROR;
1745 if (s->offset || (size & ~PAGE_MASK) || size + s->length > max) {
1746 if (__map_sg_chunk(dev, start, size, &dma->dma_address,
1747 dir, attrs, is_coherent) < 0)
1750 dma->dma_address += offset;
1751 dma->dma_length = size - offset;
1753 size = offset = s->offset;
1760 if (__map_sg_chunk(dev, start, size, &dma->dma_address, dir, attrs,
1764 dma->dma_address += offset;
1765 dma->dma_length = size - offset;
1770 for_each_sg(sg, s, count, i)
1771 __iommu_remove_mapping(dev, sg_dma_address(s), sg_dma_len(s));
1776 * arm_coherent_iommu_map_sg - map a set of SG buffers for streaming mode DMA
1777 * @dev: valid struct device pointer
1778 * @sg: list of buffers
1779 * @nents: number of buffers to map
1780 * @dir: DMA transfer direction
1782 * Map a set of i/o coherent buffers described by scatterlist in streaming
1783 * mode for DMA. The scatter gather list elements are merged together (if
1784 * possible) and tagged with the appropriate dma address and length. They are
1785 * obtained via sg_dma_{address,length}.
1787 int arm_coherent_iommu_map_sg(struct device *dev, struct scatterlist *sg,
1788 int nents, enum dma_data_direction dir, unsigned long attrs)
1790 return __iommu_map_sg(dev, sg, nents, dir, attrs, true);
1794 * arm_iommu_map_sg - map a set of SG buffers for streaming mode DMA
1795 * @dev: valid struct device pointer
1796 * @sg: list of buffers
1797 * @nents: number of buffers to map
1798 * @dir: DMA transfer direction
1800 * Map a set of buffers described by scatterlist in streaming mode for DMA.
1801 * The scatter gather list elements are merged together (if possible) and
1802 * tagged with the appropriate dma address and length. They are obtained via
1803 * sg_dma_{address,length}.
1805 int arm_iommu_map_sg(struct device *dev, struct scatterlist *sg,
1806 int nents, enum dma_data_direction dir, unsigned long attrs)
1808 return __iommu_map_sg(dev, sg, nents, dir, attrs, false);
1811 static void __iommu_unmap_sg(struct device *dev, struct scatterlist *sg,
1812 int nents, enum dma_data_direction dir,
1813 unsigned long attrs, bool is_coherent)
1815 struct scatterlist *s;
1818 for_each_sg(sg, s, nents, i) {
1820 __iommu_remove_mapping(dev, sg_dma_address(s),
1822 if (!is_coherent && (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1823 __dma_page_dev_to_cpu(sg_page(s), s->offset,
1829 * arm_coherent_iommu_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
1830 * @dev: valid struct device pointer
1831 * @sg: list of buffers
1832 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
1833 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1835 * Unmap a set of streaming mode DMA translations. Again, CPU access
1836 * rules concerning calls here are the same as for dma_unmap_single().
1838 void arm_coherent_iommu_unmap_sg(struct device *dev, struct scatterlist *sg,
1839 int nents, enum dma_data_direction dir,
1840 unsigned long attrs)
1842 __iommu_unmap_sg(dev, sg, nents, dir, attrs, true);
1846 * arm_iommu_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
1847 * @dev: valid struct device pointer
1848 * @sg: list of buffers
1849 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
1850 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1852 * Unmap a set of streaming mode DMA translations. Again, CPU access
1853 * rules concerning calls here are the same as for dma_unmap_single().
1855 void arm_iommu_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
1856 enum dma_data_direction dir,
1857 unsigned long attrs)
1859 __iommu_unmap_sg(dev, sg, nents, dir, attrs, false);
1863 * arm_iommu_sync_sg_for_cpu
1864 * @dev: valid struct device pointer
1865 * @sg: list of buffers
1866 * @nents: number of buffers to map (returned from dma_map_sg)
1867 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1869 void arm_iommu_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1870 int nents, enum dma_data_direction dir)
1872 struct scatterlist *s;
1875 for_each_sg(sg, s, nents, i)
1876 __dma_page_dev_to_cpu(sg_page(s), s->offset, s->length, dir);
1881 * arm_iommu_sync_sg_for_device
1882 * @dev: valid struct device pointer
1883 * @sg: list of buffers
1884 * @nents: number of buffers to map (returned from dma_map_sg)
1885 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
1887 void arm_iommu_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1888 int nents, enum dma_data_direction dir)
1890 struct scatterlist *s;
1893 for_each_sg(sg, s, nents, i)
1894 __dma_page_cpu_to_dev(sg_page(s), s->offset, s->length, dir);
1899 * arm_coherent_iommu_map_page
1900 * @dev: valid struct device pointer
1901 * @page: page that buffer resides in
1902 * @offset: offset into page for start of buffer
1903 * @size: size of buffer to map
1904 * @dir: DMA transfer direction
1906 * Coherent IOMMU aware version of arm_dma_map_page()
1908 static dma_addr_t arm_coherent_iommu_map_page(struct device *dev, struct page *page,
1909 unsigned long offset, size_t size, enum dma_data_direction dir,
1910 unsigned long attrs)
1912 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1913 dma_addr_t dma_addr;
1914 int ret, prot, len = PAGE_ALIGN(size + offset);
1916 dma_addr = __alloc_iova(mapping, len);
1917 if (dma_addr == ARM_MAPPING_ERROR)
1920 prot = __dma_info_to_prot(dir, attrs);
1922 ret = iommu_map(mapping->domain, dma_addr, page_to_phys(page), len, prot);
1926 return dma_addr + offset;
1928 __free_iova(mapping, dma_addr, len);
1929 return ARM_MAPPING_ERROR;
1933 * arm_iommu_map_page
1934 * @dev: valid struct device pointer
1935 * @page: page that buffer resides in
1936 * @offset: offset into page for start of buffer
1937 * @size: size of buffer to map
1938 * @dir: DMA transfer direction
1940 * IOMMU aware version of arm_dma_map_page()
1942 static dma_addr_t arm_iommu_map_page(struct device *dev, struct page *page,
1943 unsigned long offset, size_t size, enum dma_data_direction dir,
1944 unsigned long attrs)
1946 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1947 __dma_page_cpu_to_dev(page, offset, size, dir);
1949 return arm_coherent_iommu_map_page(dev, page, offset, size, dir, attrs);
1953 * arm_coherent_iommu_unmap_page
1954 * @dev: valid struct device pointer
1955 * @handle: DMA address of buffer
1956 * @size: size of buffer (same as passed to dma_map_page)
1957 * @dir: DMA transfer direction (same as passed to dma_map_page)
1959 * Coherent IOMMU aware version of arm_dma_unmap_page()
1961 static void arm_coherent_iommu_unmap_page(struct device *dev, dma_addr_t handle,
1962 size_t size, enum dma_data_direction dir, unsigned long attrs)
1964 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1965 dma_addr_t iova = handle & PAGE_MASK;
1966 int offset = handle & ~PAGE_MASK;
1967 int len = PAGE_ALIGN(size + offset);
1972 iommu_unmap(mapping->domain, iova, len);
1973 __free_iova(mapping, iova, len);
1977 * arm_iommu_unmap_page
1978 * @dev: valid struct device pointer
1979 * @handle: DMA address of buffer
1980 * @size: size of buffer (same as passed to dma_map_page)
1981 * @dir: DMA transfer direction (same as passed to dma_map_page)
1983 * IOMMU aware version of arm_dma_unmap_page()
1985 static void arm_iommu_unmap_page(struct device *dev, dma_addr_t handle,
1986 size_t size, enum dma_data_direction dir, unsigned long attrs)
1988 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
1989 dma_addr_t iova = handle & PAGE_MASK;
1990 struct page *page = phys_to_page(iommu_iova_to_phys(mapping->domain, iova));
1991 int offset = handle & ~PAGE_MASK;
1992 int len = PAGE_ALIGN(size + offset);
1997 if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
1998 __dma_page_dev_to_cpu(page, offset, size, dir);
2000 iommu_unmap(mapping->domain, iova, len);
2001 __free_iova(mapping, iova, len);
2005 * arm_iommu_map_resource - map a device resource for DMA
2006 * @dev: valid struct device pointer
2007 * @phys_addr: physical address of resource
2008 * @size: size of resource to map
2009 * @dir: DMA transfer direction
2011 static dma_addr_t arm_iommu_map_resource(struct device *dev,
2012 phys_addr_t phys_addr, size_t size,
2013 enum dma_data_direction dir, unsigned long attrs)
2015 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2016 dma_addr_t dma_addr;
2018 phys_addr_t addr = phys_addr & PAGE_MASK;
2019 unsigned int offset = phys_addr & ~PAGE_MASK;
2020 size_t len = PAGE_ALIGN(size + offset);
2022 dma_addr = __alloc_iova(mapping, len);
2023 if (dma_addr == ARM_MAPPING_ERROR)
2026 prot = __dma_info_to_prot(dir, attrs) | IOMMU_MMIO;
2028 ret = iommu_map(mapping->domain, dma_addr, addr, len, prot);
2032 return dma_addr + offset;
2034 __free_iova(mapping, dma_addr, len);
2035 return ARM_MAPPING_ERROR;
2039 * arm_iommu_unmap_resource - unmap a device DMA resource
2040 * @dev: valid struct device pointer
2041 * @dma_handle: DMA address to resource
2042 * @size: size of resource to map
2043 * @dir: DMA transfer direction
2045 static void arm_iommu_unmap_resource(struct device *dev, dma_addr_t dma_handle,
2046 size_t size, enum dma_data_direction dir,
2047 unsigned long attrs)
2049 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2050 dma_addr_t iova = dma_handle & PAGE_MASK;
2051 unsigned int offset = dma_handle & ~PAGE_MASK;
2052 size_t len = PAGE_ALIGN(size + offset);
2057 iommu_unmap(mapping->domain, iova, len);
2058 __free_iova(mapping, iova, len);
2061 static void arm_iommu_sync_single_for_cpu(struct device *dev,
2062 dma_addr_t handle, size_t size, enum dma_data_direction dir)
2064 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2065 dma_addr_t iova = handle & PAGE_MASK;
2066 struct page *page = phys_to_page(iommu_iova_to_phys(mapping->domain, iova));
2067 unsigned int offset = handle & ~PAGE_MASK;
2072 __dma_page_dev_to_cpu(page, offset, size, dir);
2075 static void arm_iommu_sync_single_for_device(struct device *dev,
2076 dma_addr_t handle, size_t size, enum dma_data_direction dir)
2078 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2079 dma_addr_t iova = handle & PAGE_MASK;
2080 struct page *page = phys_to_page(iommu_iova_to_phys(mapping->domain, iova));
2081 unsigned int offset = handle & ~PAGE_MASK;
2086 __dma_page_cpu_to_dev(page, offset, size, dir);
2089 const struct dma_map_ops iommu_ops = {
2090 .alloc = arm_iommu_alloc_attrs,
2091 .free = arm_iommu_free_attrs,
2092 .mmap = arm_iommu_mmap_attrs,
2093 .get_sgtable = arm_iommu_get_sgtable,
2095 .map_page = arm_iommu_map_page,
2096 .unmap_page = arm_iommu_unmap_page,
2097 .sync_single_for_cpu = arm_iommu_sync_single_for_cpu,
2098 .sync_single_for_device = arm_iommu_sync_single_for_device,
2100 .map_sg = arm_iommu_map_sg,
2101 .unmap_sg = arm_iommu_unmap_sg,
2102 .sync_sg_for_cpu = arm_iommu_sync_sg_for_cpu,
2103 .sync_sg_for_device = arm_iommu_sync_sg_for_device,
2105 .map_resource = arm_iommu_map_resource,
2106 .unmap_resource = arm_iommu_unmap_resource,
2108 .mapping_error = arm_dma_mapping_error,
2109 .dma_supported = arm_dma_supported,
2112 const struct dma_map_ops iommu_coherent_ops = {
2113 .alloc = arm_coherent_iommu_alloc_attrs,
2114 .free = arm_coherent_iommu_free_attrs,
2115 .mmap = arm_coherent_iommu_mmap_attrs,
2116 .get_sgtable = arm_iommu_get_sgtable,
2118 .map_page = arm_coherent_iommu_map_page,
2119 .unmap_page = arm_coherent_iommu_unmap_page,
2121 .map_sg = arm_coherent_iommu_map_sg,
2122 .unmap_sg = arm_coherent_iommu_unmap_sg,
2124 .map_resource = arm_iommu_map_resource,
2125 .unmap_resource = arm_iommu_unmap_resource,
2127 .mapping_error = arm_dma_mapping_error,
2128 .dma_supported = arm_dma_supported,
2132 * arm_iommu_create_mapping
2133 * @bus: pointer to the bus holding the client device (for IOMMU calls)
2134 * @base: start address of the valid IO address space
2135 * @size: maximum size of the valid IO address space
2137 * Creates a mapping structure which holds information about used/unused
2138 * IO address ranges, which is required to perform memory allocation and
2139 * mapping with IOMMU aware functions.
2141 * The client device need to be attached to the mapping with
2142 * arm_iommu_attach_device function.
2144 struct dma_iommu_mapping *
2145 arm_iommu_create_mapping(struct bus_type *bus, dma_addr_t base, u64 size)
2147 unsigned int bits = size >> PAGE_SHIFT;
2148 unsigned int bitmap_size = BITS_TO_LONGS(bits) * sizeof(long);
2149 struct dma_iommu_mapping *mapping;
2153 /* currently only 32-bit DMA address space is supported */
2154 if (size > DMA_BIT_MASK(32) + 1)
2155 return ERR_PTR(-ERANGE);
2158 return ERR_PTR(-EINVAL);
2160 if (bitmap_size > PAGE_SIZE) {
2161 extensions = bitmap_size / PAGE_SIZE;
2162 bitmap_size = PAGE_SIZE;
2165 mapping = kzalloc(sizeof(struct dma_iommu_mapping), GFP_KERNEL);
2169 mapping->bitmap_size = bitmap_size;
2170 mapping->bitmaps = kcalloc(extensions, sizeof(unsigned long *),
2172 if (!mapping->bitmaps)
2175 mapping->bitmaps[0] = kzalloc(bitmap_size, GFP_KERNEL);
2176 if (!mapping->bitmaps[0])
2179 mapping->nr_bitmaps = 1;
2180 mapping->extensions = extensions;
2181 mapping->base = base;
2182 mapping->bits = BITS_PER_BYTE * bitmap_size;
2184 spin_lock_init(&mapping->lock);
2186 mapping->domain = iommu_domain_alloc(bus);
2187 if (!mapping->domain)
2190 kref_init(&mapping->kref);
2193 kfree(mapping->bitmaps[0]);
2195 kfree(mapping->bitmaps);
2199 return ERR_PTR(err);
2201 EXPORT_SYMBOL_GPL(arm_iommu_create_mapping);
2203 static void release_iommu_mapping(struct kref *kref)
2206 struct dma_iommu_mapping *mapping =
2207 container_of(kref, struct dma_iommu_mapping, kref);
2209 iommu_domain_free(mapping->domain);
2210 for (i = 0; i < mapping->nr_bitmaps; i++)
2211 kfree(mapping->bitmaps[i]);
2212 kfree(mapping->bitmaps);
2216 static int extend_iommu_mapping(struct dma_iommu_mapping *mapping)
2220 if (mapping->nr_bitmaps >= mapping->extensions)
2223 next_bitmap = mapping->nr_bitmaps;
2224 mapping->bitmaps[next_bitmap] = kzalloc(mapping->bitmap_size,
2226 if (!mapping->bitmaps[next_bitmap])
2229 mapping->nr_bitmaps++;
2234 void arm_iommu_release_mapping(struct dma_iommu_mapping *mapping)
2237 kref_put(&mapping->kref, release_iommu_mapping);
2239 EXPORT_SYMBOL_GPL(arm_iommu_release_mapping);
2241 static int __arm_iommu_attach_device(struct device *dev,
2242 struct dma_iommu_mapping *mapping)
2246 err = iommu_attach_device(mapping->domain, dev);
2250 kref_get(&mapping->kref);
2251 to_dma_iommu_mapping(dev) = mapping;
2253 pr_debug("Attached IOMMU controller to %s device.\n", dev_name(dev));
2258 * arm_iommu_attach_device
2259 * @dev: valid struct device pointer
2260 * @mapping: io address space mapping structure (returned from
2261 * arm_iommu_create_mapping)
2263 * Attaches specified io address space mapping to the provided device.
2264 * This replaces the dma operations (dma_map_ops pointer) with the
2265 * IOMMU aware version.
2267 * More than one client might be attached to the same io address space
2270 int arm_iommu_attach_device(struct device *dev,
2271 struct dma_iommu_mapping *mapping)
2275 err = __arm_iommu_attach_device(dev, mapping);
2279 set_dma_ops(dev, &iommu_ops);
2282 EXPORT_SYMBOL_GPL(arm_iommu_attach_device);
2285 * arm_iommu_detach_device
2286 * @dev: valid struct device pointer
2288 * Detaches the provided device from a previously attached map.
2289 * This voids the dma operations (dma_map_ops pointer)
2291 void arm_iommu_detach_device(struct device *dev)
2293 struct dma_iommu_mapping *mapping;
2295 mapping = to_dma_iommu_mapping(dev);
2297 dev_warn(dev, "Not attached\n");
2301 iommu_detach_device(mapping->domain, dev);
2302 kref_put(&mapping->kref, release_iommu_mapping);
2303 to_dma_iommu_mapping(dev) = NULL;
2304 set_dma_ops(dev, arm_get_dma_map_ops(dev->archdata.dma_coherent));
2306 pr_debug("Detached IOMMU controller from %s device.\n", dev_name(dev));
2308 EXPORT_SYMBOL_GPL(arm_iommu_detach_device);
2310 static const struct dma_map_ops *arm_get_iommu_dma_map_ops(bool coherent)
2312 return coherent ? &iommu_coherent_ops : &iommu_ops;
2315 static bool arm_setup_iommu_dma_ops(struct device *dev, u64 dma_base, u64 size,
2316 const struct iommu_ops *iommu)
2318 struct dma_iommu_mapping *mapping;
2323 mapping = arm_iommu_create_mapping(dev->bus, dma_base, size);
2324 if (IS_ERR(mapping)) {
2325 pr_warn("Failed to create %llu-byte IOMMU mapping for device %s\n",
2326 size, dev_name(dev));
2330 if (__arm_iommu_attach_device(dev, mapping)) {
2331 pr_warn("Failed to attached device %s to IOMMU_mapping\n",
2333 arm_iommu_release_mapping(mapping);
2340 static void arm_teardown_iommu_dma_ops(struct device *dev)
2342 struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
2347 arm_iommu_detach_device(dev);
2348 arm_iommu_release_mapping(mapping);
2353 static bool arm_setup_iommu_dma_ops(struct device *dev, u64 dma_base, u64 size,
2354 const struct iommu_ops *iommu)
2359 static void arm_teardown_iommu_dma_ops(struct device *dev) { }
2361 #define arm_get_iommu_dma_map_ops arm_get_dma_map_ops
2363 #endif /* CONFIG_ARM_DMA_USE_IOMMU */
2365 void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
2366 const struct iommu_ops *iommu, bool coherent)
2368 const struct dma_map_ops *dma_ops;
2370 dev->archdata.dma_coherent = coherent;
2373 * Don't override the dma_ops if they have already been set. Ideally
2374 * this should be the only location where dma_ops are set, remove this
2375 * check when all other callers of set_dma_ops will have disappeared.
2380 if (arm_setup_iommu_dma_ops(dev, dma_base, size, iommu))
2381 dma_ops = arm_get_iommu_dma_map_ops(coherent);
2383 dma_ops = arm_get_dma_map_ops(coherent);
2385 set_dma_ops(dev, dma_ops);
2388 if (xen_initial_domain()) {
2389 dev->archdata.dev_dma_ops = dev->dma_ops;
2390 dev->dma_ops = xen_dma_ops;
2393 dev->archdata.dma_ops_setup = true;
2396 void arch_teardown_dma_ops(struct device *dev)
2398 if (!dev->archdata.dma_ops_setup)
2401 arm_teardown_iommu_dma_ops(dev);