2 * SWIOTLB-based DMA API implementation
4 * Copyright (C) 2012 ARM Ltd.
5 * Author: Catalin Marinas <catalin.marinas@arm.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program. If not, see <http://www.gnu.org/licenses/>.
20 #include <linux/gfp.h>
21 #include <linux/export.h>
22 #include <linux/slab.h>
23 #include <linux/genalloc.h>
24 #include <linux/dma-mapping.h>
25 #include <linux/dma-contiguous.h>
26 #include <linux/vmalloc.h>
27 #include <linux/swiotlb.h>
29 #include <asm/cacheflush.h>
31 struct dma_map_ops *dma_ops;
32 EXPORT_SYMBOL(dma_ops);
34 static pgprot_t __get_dma_pgprot(struct dma_attrs *attrs, pgprot_t prot,
37 if (!coherent || dma_get_attr(DMA_ATTR_WRITE_COMBINE, attrs))
38 return pgprot_writecombine(prot);
42 static struct gen_pool *atomic_pool;
44 #define DEFAULT_DMA_COHERENT_POOL_SIZE SZ_256K
45 static size_t atomic_pool_size = DEFAULT_DMA_COHERENT_POOL_SIZE;
47 static int __init early_coherent_pool(char *p)
49 atomic_pool_size = memparse(p, &p);
52 early_param("coherent_pool", early_coherent_pool);
54 static void *__alloc_from_pool(size_t size, struct page **ret_page)
60 WARN(1, "coherent pool not initialised!\n");
64 val = gen_pool_alloc(atomic_pool, size);
66 phys_addr_t phys = gen_pool_virt_to_phys(atomic_pool, val);
68 *ret_page = phys_to_page(phys);
75 static bool __in_atomic_pool(void *start, size_t size)
77 return addr_in_gen_pool(atomic_pool, (unsigned long)start, size);
80 static int __free_from_pool(void *start, size_t size)
82 if (!__in_atomic_pool(start, size))
85 gen_pool_free(atomic_pool, (unsigned long)start, size);
90 static void *__dma_alloc_coherent(struct device *dev, size_t size,
91 dma_addr_t *dma_handle, gfp_t flags,
92 struct dma_attrs *attrs)
95 WARN_ONCE(1, "Use an actual device structure for DMA allocation\n");
99 if (IS_ENABLED(CONFIG_ZONE_DMA) &&
100 dev->coherent_dma_mask <= DMA_BIT_MASK(32))
102 if (IS_ENABLED(CONFIG_DMA_CMA) && (flags & __GFP_WAIT)) {
105 size = PAGE_ALIGN(size);
106 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
111 *dma_handle = phys_to_dma(dev, page_to_phys(page));
112 return page_address(page);
114 return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
118 static void __dma_free_coherent(struct device *dev, size_t size,
119 void *vaddr, dma_addr_t dma_handle,
120 struct dma_attrs *attrs)
123 phys_addr_t paddr = dma_to_phys(dev, dma_handle);
126 WARN_ONCE(1, "Use an actual device structure for DMA allocation\n");
130 freed = dma_release_from_contiguous(dev,
134 swiotlb_free_coherent(dev, size, vaddr, dma_handle);
137 static void *__dma_alloc(struct device *dev, size_t size,
138 dma_addr_t *dma_handle, gfp_t flags,
139 struct dma_attrs *attrs)
142 void *ptr, *coherent_ptr;
143 bool coherent = is_device_dma_coherent(dev);
145 size = PAGE_ALIGN(size);
147 if (!coherent && !(flags & __GFP_WAIT)) {
148 struct page *page = NULL;
149 void *addr = __alloc_from_pool(size, &page);
152 *dma_handle = phys_to_dma(dev, page_to_phys(page));
157 ptr = __dma_alloc_coherent(dev, size, dma_handle, flags, attrs);
161 /* no need for non-cacheable mapping if coherent */
165 /* remove any dirty cache lines on the kernel alias */
166 __dma_flush_range(ptr, ptr + size);
168 /* create a coherent mapping */
169 page = virt_to_page(ptr);
170 coherent_ptr = dma_common_contiguous_remap(page, size, VM_USERMAP,
171 __get_dma_pgprot(attrs,
172 __pgprot(PROT_NORMAL_NC), false),
180 __dma_free_coherent(dev, size, ptr, *dma_handle, attrs);
182 *dma_handle = DMA_ERROR_CODE;
186 static void __dma_free(struct device *dev, size_t size,
187 void *vaddr, dma_addr_t dma_handle,
188 struct dma_attrs *attrs)
190 void *swiotlb_addr = phys_to_virt(dma_to_phys(dev, dma_handle));
192 if (!is_device_dma_coherent(dev)) {
193 if (__free_from_pool(vaddr, size))
197 __dma_free_coherent(dev, size, swiotlb_addr, dma_handle, attrs);
200 static dma_addr_t __swiotlb_map_page(struct device *dev, struct page *page,
201 unsigned long offset, size_t size,
202 enum dma_data_direction dir,
203 struct dma_attrs *attrs)
207 dev_addr = swiotlb_map_page(dev, page, offset, size, dir, attrs);
208 if (!is_device_dma_coherent(dev))
209 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
215 static void __swiotlb_unmap_page(struct device *dev, dma_addr_t dev_addr,
216 size_t size, enum dma_data_direction dir,
217 struct dma_attrs *attrs)
219 if (!is_device_dma_coherent(dev))
220 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
221 swiotlb_unmap_page(dev, dev_addr, size, dir, attrs);
224 static int __swiotlb_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
225 int nelems, enum dma_data_direction dir,
226 struct dma_attrs *attrs)
228 struct scatterlist *sg;
231 ret = swiotlb_map_sg_attrs(dev, sgl, nelems, dir, attrs);
232 if (!is_device_dma_coherent(dev))
233 for_each_sg(sgl, sg, ret, i)
234 __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
240 static void __swiotlb_unmap_sg_attrs(struct device *dev,
241 struct scatterlist *sgl, int nelems,
242 enum dma_data_direction dir,
243 struct dma_attrs *attrs)
245 struct scatterlist *sg;
248 if (!is_device_dma_coherent(dev))
249 for_each_sg(sgl, sg, nelems, i)
250 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
252 swiotlb_unmap_sg_attrs(dev, sgl, nelems, dir, attrs);
255 static void __swiotlb_sync_single_for_cpu(struct device *dev,
256 dma_addr_t dev_addr, size_t size,
257 enum dma_data_direction dir)
259 if (!is_device_dma_coherent(dev))
260 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
261 swiotlb_sync_single_for_cpu(dev, dev_addr, size, dir);
264 static void __swiotlb_sync_single_for_device(struct device *dev,
265 dma_addr_t dev_addr, size_t size,
266 enum dma_data_direction dir)
268 swiotlb_sync_single_for_device(dev, dev_addr, size, dir);
269 if (!is_device_dma_coherent(dev))
270 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
273 static void __swiotlb_sync_sg_for_cpu(struct device *dev,
274 struct scatterlist *sgl, int nelems,
275 enum dma_data_direction dir)
277 struct scatterlist *sg;
280 if (!is_device_dma_coherent(dev))
281 for_each_sg(sgl, sg, nelems, i)
282 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
284 swiotlb_sync_sg_for_cpu(dev, sgl, nelems, dir);
287 static void __swiotlb_sync_sg_for_device(struct device *dev,
288 struct scatterlist *sgl, int nelems,
289 enum dma_data_direction dir)
291 struct scatterlist *sg;
294 swiotlb_sync_sg_for_device(dev, sgl, nelems, dir);
295 if (!is_device_dma_coherent(dev))
296 for_each_sg(sgl, sg, nelems, i)
297 __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
301 /* vma->vm_page_prot must be set appropriately before calling this function */
302 static int __dma_common_mmap(struct device *dev, struct vm_area_struct *vma,
303 void *cpu_addr, dma_addr_t dma_addr, size_t size)
306 unsigned long nr_vma_pages = (vma->vm_end - vma->vm_start) >>
308 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
309 unsigned long pfn = dma_to_phys(dev, dma_addr) >> PAGE_SHIFT;
310 unsigned long off = vma->vm_pgoff;
312 if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
315 if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) {
316 ret = remap_pfn_range(vma, vma->vm_start,
318 vma->vm_end - vma->vm_start,
325 static int __swiotlb_mmap(struct device *dev,
326 struct vm_area_struct *vma,
327 void *cpu_addr, dma_addr_t dma_addr, size_t size,
328 struct dma_attrs *attrs)
330 vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot,
331 is_device_dma_coherent(dev));
332 return __dma_common_mmap(dev, vma, cpu_addr, dma_addr, size);
335 static struct dma_map_ops swiotlb_dma_ops = {
336 .alloc = __dma_alloc,
338 .mmap = __swiotlb_mmap,
339 .map_page = __swiotlb_map_page,
340 .unmap_page = __swiotlb_unmap_page,
341 .map_sg = __swiotlb_map_sg_attrs,
342 .unmap_sg = __swiotlb_unmap_sg_attrs,
343 .sync_single_for_cpu = __swiotlb_sync_single_for_cpu,
344 .sync_single_for_device = __swiotlb_sync_single_for_device,
345 .sync_sg_for_cpu = __swiotlb_sync_sg_for_cpu,
346 .sync_sg_for_device = __swiotlb_sync_sg_for_device,
347 .dma_supported = swiotlb_dma_supported,
348 .mapping_error = swiotlb_dma_mapping_error,
351 extern int swiotlb_late_init_with_default_size(size_t default_size);
353 static int __init atomic_pool_init(void)
355 pgprot_t prot = __pgprot(PROT_NORMAL_NC);
356 unsigned long nr_pages = atomic_pool_size >> PAGE_SHIFT;
359 unsigned int pool_size_order = get_order(atomic_pool_size);
361 if (dev_get_cma_area(NULL))
362 page = dma_alloc_from_contiguous(NULL, nr_pages,
365 page = alloc_pages(GFP_DMA, pool_size_order);
369 void *page_addr = page_address(page);
371 memset(page_addr, 0, atomic_pool_size);
372 __dma_flush_range(page_addr, page_addr + atomic_pool_size);
374 atomic_pool = gen_pool_create(PAGE_SHIFT, -1);
378 addr = dma_common_contiguous_remap(page, atomic_pool_size,
379 VM_USERMAP, prot, atomic_pool_init);
382 goto destroy_genpool;
384 ret = gen_pool_add_virt(atomic_pool, (unsigned long)addr,
386 atomic_pool_size, -1);
390 gen_pool_set_algo(atomic_pool,
391 gen_pool_first_fit_order_align,
394 pr_info("DMA: preallocated %zu KiB pool for atomic allocations\n",
395 atomic_pool_size / 1024);
401 dma_common_free_remap(addr, atomic_pool_size, VM_USERMAP);
403 gen_pool_destroy(atomic_pool);
406 if (!dma_release_from_contiguous(NULL, page, nr_pages))
407 __free_pages(page, pool_size_order);
409 pr_err("DMA: failed to allocate %zu KiB pool for atomic coherent allocation\n",
410 atomic_pool_size / 1024);
414 static int __init swiotlb_late_init(void)
416 size_t swiotlb_size = min(SZ_64M, MAX_ORDER_NR_PAGES << PAGE_SHIFT);
418 dma_ops = &swiotlb_dma_ops;
420 return swiotlb_late_init_with_default_size(swiotlb_size);
423 static int __init arm64_dma_init(void)
427 ret |= swiotlb_late_init();
428 ret |= atomic_pool_init();
432 arch_initcall(arm64_dma_init);
434 #define PREALLOC_DMA_DEBUG_ENTRIES 4096
436 static int __init dma_debug_do_init(void)
438 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
441 fs_initcall(dma_debug_do_init);